IAES Inter national J our nal of Artificial Intelligence (IJ-AI) V ol. 15, No. 4, August 2026, pp. 3761 3769 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i4.pp3761-3769 r 3761 Deep lear ning f or categorizing micr osatellite stability in color ectal cancer Sofyan El Idrissi, Y assine Drider , Ikram Ben Abdel Ouahab, Mohammed Bouhorma, F atiha El Ouaai C3S Laboratory , F aculty of Sciences and T echniques of T angier , Uni v ersity Abdelmalek Essaadi, T etouan, Morocco Article Inf o Article history: Recei v ed Aug 3, 2024 Re vised Jun 10, 2026 Accepted Jul 9, 2026 K eyw ords: Biomark ers Colorectal cancer Con v olutional neural netw orks Deep learning Digital P athology Histopathology Microsatellite instability ABSTRA CT Cancer remains a significant global health challenge, wit h its incidence rising steadily in rec ent decades. In colorectal cancer (CRC), microsatellite instability (MSI), and microsatellite stability (MSS) are important biomark ers that influence treatment decisions and patient outcomes. Accurate MSI classification is critical b ut traditional methods can be costly and time-consuming. This study e xplores the potential of deep learning to clas sify MSI and MSS in CRC. A lar ge dataset of CRC patients with confirmed MSI and MSS status w as utilized, obtained through standard testing images. Deep learning models were applied to histopathological images, analyzing tissue features from digital slides. Con v olutional neural netw ork (CNN) and residual netw ork (ResNet)-18 models demonstrated hi gh accurac y in distinguishing between MSI and MS S CRCs. The best-performing model, which inte grated genomic and histopathological data, achie v ed an area under the curv e (A UC) recei v er operating characteristic (R OC) of 0.85, indicating strong discrimination capability . The findi ngs suggest that deep learning could be a v aluable tool for clinical decision-making and personalized medicine in CRC. This is an open access article under the CC BY -SA license . Corresponding A uthor: Sofyan El Idrissi C3S Laboratory , F aculty of Sciences and T echniques of T angier , Uni v ersity Abdelmalek Essaadi T etouan, Morocco Email: idrissisofyan38@gmail.com 1. INTR ODUCTION Colorectal cancer (CRC) remains a for midable global health challenge, currently rank ed as the t hird most pre v alent malignanc y and the second leading cause of cancer -related mortality w orldwide [1]. Originating within the epithelial cells of the colon or rectum, CRC typically arises from a sequential accumulation of genetic and epigenetic alterations that dri v e uncontroll ed cellular proliferation [2]. These pathological changes often manifest initially as benign polyps which, o v er time, may transform into malignant neoplasms capable of in v ading adjacent tissues or metastasizing to dist ant or g ans via the lymphatic and circulatory systems, thereby se v erely compromising patient prognosis [3]. The biological heterogeneity of CRC necessitates precise molecular characterization to guide clinical management. While modifiable lifestyle f actors, including sedentary beha vior and diets high in processed meats, interact with non-modifiable risks such as aging to influence disease incidence [4], contemporary oncology increasingly relies on genomic biomark ers to na vig ate the substantial economic and ph ysiological b urdens of long-term treatment [5]. A fundamental molecular di stinction in CRC is the status of microsatellite stability ( MSS). Approximately 15% of cases e xhibit microsatellite instability (MSI), a condition resulting from a deficient DN A mismatch repair (dMMR) system [6]. MSI-H tumors are characterized by a high frequenc y of mutations J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
3762 r ISSN: 2252-8938 in short, repetiti v e DN A sequences, leading to a “h ypermutated” genomic profile. Histologically , these tumors often present with poor dif ferentiation and a dense infiltration of tumor -infiltrating lymphoc ytes (TILs) [7]. This high tumor mutational b urden (TMB) generates immunogenic neoantigens that render MSI-H patients e xceptionally responsi v e to immune checkpoint inhibitors (ICIs), s uch as pembrolizumab and ni v olumab, despite their frequent resistance to con v entional chemotherap y [8], [9]. Con v ersely , MSS tumors often described as “immune-cold”, e xhibit lo w mutational profiles and sparse T -cell infiltration, making them intrinsically resistant to single-agent immunotherap y [10]. These tumors are typically dri v en by chromosomal instability , in v olving mutations in oncogenes such as KRAS and tumor suppressors lik e TP53 [11]. F or these patients, clinical strate gies focus on c ytotoxic chemotherap y combined with tar geted therapies directed at epidermal gro wth f act or receptor (EGFR) or v ascular endothelial gro wth f actor (VEGF) pathw ays [12], [13]. While the distinction between MSI and MSS is critical for optimizing survi v al and treatment cost-ef fecti v eness [14], gold-standard diagnostic methods lik e immunohistochemistry (IHC) or polymerase chain reaction (PCR) are often resource-intensi v e, time-consuming, and require specialized pathology e xpertise. R ecent breakthroughs in digital pathology suggest that deep learning can potentially predict these molecular biomark ers directly from standard hematoxylin and eosin (H&E) stained slides [15]. Despite these adv ancements, a significant g ap remains in the literature re g arding the de v elopment of lo w-cost, automated MSI/MSS classifiers that are rigorously v alidated on lar ge-scale cohorts such as the Cancer Genome Atlas (TCGA). This study utilizes deep learning o v er c o n v entional statistical methods due to its superior ca p a city for hierarchical feature e xtraction, enabling the detect ion of subtle morphological patterns in histology images that e v ade human perception. The primary objecti v es of this research ar e to design a customized con v olutional neural netw ork (CNN) architecture optimized for CRC screening, conduct a comparati v e performance e v aluation ag ainst the established residual netw ork (ResNet)-18 benchmark, and assess the clinical feasibility of these models as automated decision-support tools for oncologists. 2. RELA TED W ORKS Although clinically defined by its anatomical origin, CRC is a highly genetically heterogeneous disease. This molecular di v ersity influences tumor beha vior , treatment response, and patient outcomes. Despite this v ariability , s tandard chemotherap y protocols, typically FOLFO X or FOLFIRI combined with tar geted agents are often applied uniformly , with limited consideration of indi vidual pharmacogenetic v ariations and molecular characteris tics. Adjuv ant management remains primarily dictated by tumor stage. The pi v otal MOSAIC study established that the addit ion of oxaliplatin to 5-fluorouracil (5-FU) significantly reduces recurrence risk in stage II and III disease [16], demonstrating that early-stage patients can deri v e considerable curati v e benefit. W ithin this frame w ork, MSI status has emer ged as a critical predicti v e mark er for chemotherap y ef ficac y . While the oxaliplatin/5-FU combination is the established adjuv ant standard for resected stage III cancers, the MSI phenotype is historically associated with intrinsic resistance to fluorop yrimidine monotherap y . Pooled analyses of randomized trials confirm t hat while MSI patients often e xhibit prolonged survi v al, the y deri v e no benefit from adjuv ant 5-FU alone; in f act, potential deleterious ef fects ha v e been observ ed in stage II disease. Ho we v er , the inte gration of oxaliplatin has been sho wn to o v ercome this resistance, restoring therapeutic ef ficac y specifically for stage III tumors [17]-[19]. Furthermore, findings from the IDEA international collaborati v e study ha v e refined the duration and modalities of t hese oxaliplatin-based re gimens for stage III patients [20]. In the metastatic setting, distinguishing between MSS and MSI phenotypes is a fundamental c riterion for therapeutic orientation. P atients with metastatic MSI cancer are no w primary candidates for immune checkpoint blockade (ICB) using anti-PD-1 or anti-CTLA-4 antibodies [21]. This immunotherap y protocol is highly ef fecti v e in MSI-H CRC b ut remains lar gely inef fecti v e in MSS cases [22]. Consequently , clinical guidelines, such as the French Na tional Digesti v e Cancer Thesaurus (TNCD), recommend adjuv ant chemotherap y (fluorop yrimidine + oxaliplatin) for stage III MSI tumors, while generally omitting adjuv ant treatment for stage II MSI tumors unless high-risk features, such as pT4b status, are present [10]. While the MSI/MSS phenotype can be identified through con v entional IHC or genetic analysis [23], these tests are often e xpensi v e or una v ailable in resource-constrained re gions, lea ding to suboptimal screening. Notably , MSI-H CRCs e xhibit distinct histomorphological features compared to MSS counterparts [24]. This morphological di v er gence has pa v ed the w ay for deep learning algorithms to analyze histopathological images for survi v al prediction [25], metastasis detection [26], and precise histological grading [27], [28]. Int J Artif Intell, V ol. 15, No. 4, August 2026: 3761–3769 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 r 3763 Recent research increasingly demonstrates that deep learning model s can bridge the g ap be tween histological patterns and underlying genomic mutations [29]. F or instance, Kather et al . [30] pro v ed the ef fecti v eness of predicting MSI status directly from H&E-stained slides in g astrointestinal cancers, a finding supported by the interacti v e HistomicsML system [31] and subsequent studies [32]. Furthermore, K e et al . [33] utilized high-performance computing (NVIDIA T esla V100) to achie v e an MSI prediction accurac y of 87%, while Lee et al . [24] estimated that up to 40% of CRC slides could be ef fecti v ely screened via deep learning without molecular testing. Adv anced models such as MSINet ha v e achie v ed an area under the curv e - recei v er o pe rating characteristic (A UC-R OC) of 0.931 [34], and artificial intelligence (AI)-based stratification has sho wn superior results in cross- v a lidation with an area under the recei v er operating characteristic (A UR OC) of 0.86 [35]. This progress is mirrored in consensus molecular subtype (CMS) classification [36] and the inte gration of se gment an ything model (SAM) with deep learning, which has further refined the se gmentation and classification of MSI, un v eiling its full clinical significance in contemporary CRC management [37]. 3. METHODS 3.1. Dataset TCGA initiati v e pro vides comprehensi v e molecular portraits of v arious malignancies, notably including the colon adenocarcinoma (CO AD) cohort. These multi-omic datasets enable the prediction of MSI status by inte grating molecular profiles, dif ferential gene e xpression, and DN A meth ylation patterns [38]. F or this study , the dataset consists of high-resolution digital pathology slides processed into standardized JPEG tiles, capturing the distinct morphological features associated with MSI-MUT and MSS phenotypes. The data preprocessing pipeline w as engineered to structure and partition these image tiles into three distinct subsets: training (60%), v alidation (25%), and testing (15%). T o miti g ate selection bi as and ensure statistical representati v eness, the allocation process utilizes a randomized shuf fling of filenames prior to distrib ution. The dataset e xhibits a class distrib ution of 39.02% for MSI-MUT JPEG and 60.98% for MSS JPEG, reflecting the inherent pre v alence v ariance found in clinical populations. T o maintain data inte grity during training, these subsets are processed through the torchvision.datasets. ImageF older class, with images resized and normalized to align with the input requirements of the deep learning architectures. Figure 1 illustrates representati v e sample images e xtracted from the dataset, highlighting the te xtural and cellular di v ersities between the tw o molecular subtypes. This figure displaying representati v e patches for both MSI and MSS classes. Figure 1. Sample histological images from the dataset 3.2. Deep lear ning models CNNs stand tall as a pi v otal architecture, particularly in tasks related to image recogni tion, classification, and e v en in certain aspects of natural language processing. The ef ficienc y of CNNs can be attrib uted to their core elements, each playing a distinct role in information e xtraction and abstraction within the netw ork. Here, this study use the fundamental components of CNNs: input dimension, con v olution layer , pooling layer , fully connected layer , and output. The complete architecture of the proposed CNN model is illustrated in Figure 2. ResNet-18 is a pi v otal CNN architecture that addresses the v anishing gradient problem through the inte gration of residual learning [39]. By utilizing shortcut connections, or skip connections, the netw ork ef fecti v ely performs identity mapping, allo wing gradients to bypass intermediate layers and f acilitating the optimization of deeper s tructures. The ResNet-18 v ariant specific ally emplo ys a balanced configuration of 18 deep la y e rs, comprising an initial 7 7 con v olution follo wed by four stages of residual blocks, each containing tw o 3 3 con v olutional layers [40], [41]. This architecture terminates with a global a v erage pooling layer and a fully connected softmax classifier . While deeper iterations lik e ResNet-50 or ResNet-101 of fer higher capacity , ResNet-18 remains a standard benchmark in computer vision due to its Deep learning for cate gorizing micr osatellite stability in color ectal cancer (Sofyan El Idrissi) Evaluation Warning : The document was created with Spire.PDF for Python.
3764 r ISSN: 2252-8938 optimal tra de-of f between representat i v e po wer and computational ef ficienc y , making it particula rly suitable for resource-constrained en vironments and real-time inference tasks. The detailed structure of the pre-trained ResNet-18 model is illustrated in Figure 3. Figure 2. Schematic diagram of the CNN architecture used for MSI/MSS classification Figure 3. Architecture of the pre-trained ResNet-18 model [39] F or the e xperimental implementation, the torchvision library in Python is le v eraged to preprocess and load the TCGA dataset. Specifically , an or g anized data structure is established using the ImageF older class, while the T oT ensor transformation standardizes the image inputs. T o ensure a balanced representation across all cate gories, class distrib utions are rigorously analyzed by it erating o v er class indices and quantifying image counts for each respecti v e label. The core architecture utilizes the ResNet-18 model tailored for the TCGA data. During e xperimentation, e xtensi v e h yperparameter tuning is performed to optimize model fit; v ariations include adjustments to the number and type of layers, k ernel sizes (ranging from 3 3 to 7 7 ), padding, and pooling strate gies (max and a v erage pooling). Furthermore, the impact of ReLU acti v ation, v arying learning rates ( 0 : 001 to 0 : 1 ), and dropout rates ( 0 : 2 to 0 : 5 ) is e xplored alongside se v eral optimizers, including stochastic gradient descent (SGD), adapti v e moment estimation (Adam), and root mean square propag ation (RMSprop) and re gularization techniques ( L 1 and L 2 ). The TCGA database is partitioned into 60% for training, 25% for v alidation, and 15% f o r testing. Performance is quantified using a comprehensi v e suite of e v aluation metrics, including accurac y , precision, recall, F1-score, and the A UC. Additionally , both standard and normalized confusion matrices are analyzed to pro vide a granular assessment of classific ation performance and to f acilitate comparati v e benchmarking with e xisting literature. Computational e xperiments are conduct ed on a professional hardw are en vironment featuring an NVIDIA Quadro T2000 mobile GPU with 4 GB GDDR5 memory . This setup is augmented by Google Colab Pro sessions, pro viding access to cloud-based T4 GPU resources for more intensi v e processing requirements. 4. RESUL TS AND DISCUSSION The primary objecti v e of this research w as to e v aluate the ef ficac y of deep learning architectures in accurately discriminating between MSI and MSS molecular subtypes in CRC histology . The e xperimental frame w ork compared the performance of a transfer -learning-based ResNet-18 benchmark ag ainst a proposed Int J Artif Intell, V ol. 15, No. 4, August 2026: 3761–3769 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 r 3765 custom CNN architecture specifically optimized for high-throughput histopathological tile analysis. The empirical results, as detailed in T able 1, demonstrate that the proposed CNN significantly outperforms the ResNet-18 model across all e v aluated statistical dimensions. Specifically , the custom archit ecture achie v ed an accurac y of 87.51% and an A UC-R OC of 0.8553, representing a substantial impro v ement o v er the ResNet-18 baseline, which yielded an accurac y of 69.74% a nd an A UC of 0.6751. Furthermore, the F1-score of 0.9023 achie v ed by the proposed model confirms its rob ust capabi lity to maintain predicti v e equilibrium, whereas ResNet-18 e xhibited a mark ed susce p t ibility to the majority class bias inherent in unbalanced histopathological datasets. T able 1. Ev aluation metrics of ResNet-18 and proposed CNN models Metric ResNet-18 Proposed CNN Confusion matrix 29280 21846 17695 61886 39187 12012 4395 75838 A UC score 0.6751 0.8553 Accurac y 0.6974 0.8751 Precision 0.7390 0.8632 Recall 0.7776 0.9452 F1-score 0.7578 0.9023 A granular analysis of the confusion matrices, illustrated in Figure 4, re v eals critical insights into the clinical reliabilit y and applicability of these models. While ResNet-18 successfully identified 29,280 true positi v e MSI samples, it w as characterized by a high f alse-ne g ati v e rate, misclassifying 17,695 MSI cases as MSS. From a clinical perspecti v e, such omissions a re highly detrimental , as the y could pote ntially depri v e patients of life-sa ving immunot herap y . In contrast, the proposed CNN model drastically mitig ated these errors by correctly identifying 39,187 MSI samples and reducing f alse ne g ati v es to 4,395. This refinement is most e vident in the recall (sensiti vity) metric, which sur ged from 77.76% with ResNet-18 to 94.52% with the proposed architecture. Gi v en that high sensiti vity is a non-ne gotiable prerequisite for diagnostic screening tools, these results suggest that the proposed CNN architecture of fers a more viable pathw ay for assisting pathologists in identifying patients who w ould benefit from immune checkpoint inhibitors, striking a superior balance between representati v e po wer and computational ef ficienc y . Figure 4. Normalized confusion matrix for MSI/MSS classification, illustrating the prediction accurac y across both classes T o conte xtualize the findings within the contemporary research landscape, this study conducted a comparati v e analysis with prominent studies published between 2019 a nd 2024. The results align closely with the foundational e vidence pro vided by Kather et al . [30], who esta blished the baseline feasibility of deep learning for MSI prediction with an A UC of approximately 0.84. Notably , the proposed CNN slightly surpasses this benchmark (A UC 0.855), reinforcing the premise that streamlined, purpose-b uilt architectures can match or e xceed the performance of massi v e pre-trained netw orks when optimized for specific Deep learning for cate gorizing micr osatellite stability in color ectal cancer (Sofyan El Idrissi) Evaluation Warning : The document was created with Spire.PDF for Python.
3766 r ISSN: 2252-8938 histopathological tasks. Recent architect ural strate gies ha v e further di v ersified this field. F or instance, while Y amashita et al . [34] demonstrated the ef ficac y of MobileNetV2 and ResNet v ariants, the study highlights that a custom CNN can achie v e superior sensiti vity without the prohibiti v e computational o v erhead associated with deeper netw orks. Similarly , Du et al . [42] introduced a dual-scal e information approach to enhance predicti v e precision. While their multi-scale methodology of fers high granularity , our patch-based CNN achie v es a highly competiti v e F1-score (0.90) through a significantly more ef ficient implementation pipeline. Furthermore, the methodology of fers a distinct contrast to the c omple x classification techniques described in recent literature. The findings di v er ge from the tw o-stage frame w ork presented by Lee et al . [43], which necessitates a primary tumor se gmentation step follo wed by an Inception-ResNet-V2 classifier . Although Lee et al . [43] reported impro v ed F1-scores at 20 magnification, the model maintains consistent performance across the dataset without the prerequisite of computationally e xpensi v e pre-se gmentation. Finally , compared to the kno wledge distillation model proposed by K e et al . [44], which focuses on intricate patch-le v el sorting, the end-to-end CNN approach f acilitates a more direct and seamless clinical w orkflo w for slide cate gorization. 5. CONCLUSION Ov erall, this study places a particular emphasis on CRC, a leading cause of cancer -related morbidity and mortality , to delv e into the epidemiological patterns, risk f actors, and adv ancements in diagnosis and treatment. By e xploring the potential of deep learning models for cate gorizing MSI and MSS in CRC, this study demonstrated the capability of deep learning models to accurately discriminate between MSI and MSS in the case of CRC. This study has limitations, pri marily the class imbalance (MSS predominance) and the reliance solely on the TCGA dataset. Consequently , e xternal v alidation on independent cohorts is required to rule out site-specific biases and confirm generalizability . Future w ork will also incorporate e xplainability techniques (e.g., gradient-weighted class acti v ation mapping (Grad-CAM)) to visualize the histological features dri ving the model’ s predictions, thereby enhancing clinical trust. Further research is w arranted to v alidate actual findings in lar ge, prospecti v e clinical trials and address potential challenges. Inte grating interpretable models and addressing potential biases are crucial ne xt steps. The future w orks include also: addressing potential biases and enhancing model interpretability are essential for ethical considerations. Exploration of alternati v e deep learning architectures and data augmentation techniques could further impro v e performance. Ne v ertheless, this study pro vides compelling e vidence for the t ransformati v e potential of deep learning in enhancing CRC diagnosis and management, ultimately contrib uting to impro v ed patient outcomes. FUNDING INFORMA TION This research recei v ed no specific grant from an y funding agenc y in the public, commercial, or not-for -profit sectors. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Sofyan El Idrissi X X X X X X X X Y assine Drider X X X X X X X Ikram Ben Abdel Ouahab X X X X X X X X X Mohammed Bouhorma X X X X X X F atiha El Ouaai X X X X X C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting Int J Artif Intell, V ol. 15, No. 4, August 2026: 3761–3769 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 r 3767 CONFLICT OF INTEREST ST A TEMENT Authors state no conflict of interest. D A T A A V AILABILITY The data that support the findings of this study are a v ailable upon reasonable request to the corresponding author , [SEI]. REFERENCES [1] H. Sung et al. , “Global Cancer Statistics 2020: GLOBOCAN estimates of incidence and mortality w orldwide for 36 cancers in 185 countries, ” CA: A Cancer J ournal for Clinicians , v ol. 71, no. 3, pp. 209–249, May 2021, doi: 10.3322/caac.21660. [2] E. R. Fearon and B. V ogelstein, “ A genetic model for colorectal tumorigenesis, ” Cell , v ol. 61, no. 5, pp. 759–767, Jun. 1990, doi: 10.1016/0092-8674(90)90186-I. [3] G. P . Gupta and J. Massagu ´ e, “Cancer metastasis: b uilding a frame w ork, ” Cell , v ol. 127, no. 4, pp. 679–695, No v . 2006, doi: 10.1016/j.cell.2006.11.001. [4] M. J. Thun et al., “ Alcohol consumption and mortality among middle-aged and elderly U.S. adults, ” Ne w England J ournal of Medicine , v ol. 337, no. 24, pp. 1705–1714, Dec. 1997, doi: 10.1056/NEJM199712113372401. [5] P . Ra wla, T . Sunkara, and A. Barsouk, “Epidemiology of colorectal cancer: incidence, mortality , survi v al, and risk f actors, ” Gastr oenter olo gy Re vie w , v ol. 14, no. 2, pp. 89–103, 2019, doi: 10.5114/pg.2018.81072. [6] C. R. Boland and A. Goel, “Microsatellite instability in colorectal cancer , ” Gastr oenter olo gy , v ol. 138, no. 6, pp. 2073-2087.e3, May 2010, doi: 10.1053/j.g astro.2009.12.064. [7] T . Andr ´ e et al. , “Pembrolizumab in microsatellite-instability–high adv anced colorectal cancer , ” Ne w England J ournal of Medicine , v ol. 383, no. 23, pp. 2207–2218, Dec. 2020, doi: 10.1056/NEJMoa2017699. [8] D. T . Le et al. , “PD-1 blockade in tumors with mismatch-repair deficienc y , ” Ne w England J ournal of Medicine , v ol. 372, no. 26, pp. 2509–2520, Jun. 2015, doi: 10.1056/NEJMoa1500596. [9] A. B. Schrock et al. , “T umor mutational b urden is predicti v e of response to immune checkpoint inhibitors i n MSI-high metastatic colorectal cancer , ” Annals of Oncolo gy , v ol. 30, no. 7, pp. 1096–1103, Jul. 2019, doi: 10.1093/annonc/mdz134. [10] K. Ganesh et al. , “Immunotherap y in colorectal cancer: rationale, challenges and potential, ” Natur e Re vie ws Gastr oenter olo gy & Hepatolo gy , v ol. 16, no. 6, pp. 361–375, Jun. 2019, doi: 10.1038/s41575-019-0126-x. [11] D. M. Muzn y et al. , “Comprehensi v e molecular characterization of human colon and rectal cancer , ” Natur e , v ol. 487, no. 7407, pp. 330–337, Jul. 2012, doi: 10.1038/nature11252. [12] E. V . Cutsem et al. , “ESMO consensus guidelines for the management of patients with metastatic colorectal cancer , ” Annals of Oncolo gy , v ol. 27, no. 8, pp. 1386–1422, Aug. 2016, doi: 10.1093/annonc/mdw235. [13] A. Cerv antes et al. , “Metastatic colorectal cancer: ESMO clinical practice guideline for diagnosis, treatment and follo w-up, ” Annals of Oncolo gy , v ol. 34, no. 1, pp. 10–32, Jan. 2023, doi: 10.1016/j.annonc.2022.10.003. [14] D. J. Sar gent et al. , “Defecti v e mismatch repair as a predicti v e mark er for lack of ef ficac y of fluorouracil-based adjuv ant therap y in colon cancer , ” J ournal of Clinical Oncolo gy , v ol. 28, no. 20, pp. 3219–3226, Jul. 2010, doi: 10.1200/JCO.2009.27.1825. [15] M . J. Ov erman et al. , “Ni v olumab in pa tients with metastatic DN A m ismatch repair -deficient or microsatellite instability-hi gh colorectal cancer (CheckMate 142): an open-label, multice ntre, phas e 2 study , ” The Lancet Oncolo gy , v ol. 18, no. 9, pp. 1182–1191, Sep. 2017, doi: 10.1016/S1470-2045(17)30422-9. [16] T . Andr ´ e et al. , “Oxaliplatin, fluorouracil, and leuco v orin as adjuv ant treatment for colon cancer , ” Ne w England J ournal of Medicine , v ol. 350, no. 23, pp. 2343–2351, Jun. 2004, doi: 10.1056/NEJMoa032709. [17] A . Zaanan et al . , “Chemotherap y of adv anced small-bo wel adenoc arcinoma: a multicenter A GEO study , ” Annals of Oncolo gy , v ol. 21, no. 9, pp. 1786–1793, Sep. 2010, doi: 10.1093/annonc/mdq038. [18] D . T ougeron et al. , “Ef ficac y of adjuv ant chemotherap y in colon cancer with microsatellite instability: a lar ge multicenter A GEO study , ” J ournal of the National Cancer Institute , v ol. 108, no. 7, 2016, doi: 10.1093/jnci/djv438. [19] D. F ink et al. , “The role of DN A mismatch repair in platinum drug resistance, ” Cancer Resear c h , v ol. 56, no. 21, pp. 4881–4886, No v . 1996. [20] A. B. Benson et al. , “Colon cancer , v ersion 2.2024, NCCN clinical practice guidelines in oncology , ” J ournal of the National Compr ehensive Cancer Network , v ol. 21, 2024, doi: 10.6004/jnccn.2024.0029. [21] K. Li, H. Luo, L. Huang, H. Luo, and X. Zhu, “Microsatellite ins tability: a re vie w of what the oncologist should kno w , ” Cancer Cell International , v ol. 20, 2020, doi: 10.1186/s12935-019-1091-8. [22] V . A. D. W e ger et al. , “Clinical ef fects of adjuv ant acti v e specific immunotherap y dif fer between patients with microsatellite-stable and microsatellite-instable colon cancer , ” Clinical Cancer Resear c h , v ol. 18, no. 3, pp. 882–889, Feb . 2012, doi: 10.1158/1078-0432.CCR-11-1716. [23] J. N. Kather , N. Halama, and D. Jae ger , “Genomics and emer ging biomark ers for immunotherap y of colorectal cancer , ” Seminar s in Cancer Biolo gy , v ol. 52, pp. 189–197, Oct. 2018, doi: 10.1016/j.semcancer .2018.02.010. [24] S. H. Lee, I. H. Song, and H. Jang, “Feasibility of deep learning-based fully automated classification of microsatellite instability in tissue slides of colorectal cancer , ” International J ournal of Cancer , v ol. 149, no. 3, pp. 728–740, Aug. 2021, doi: 10.1002/ijc.33599. [25] J. N. Kather et al. , “Predicting survi v al from colorectal cancer histology slides using deep learning: a retrospecti v e multicenter study , ” PLoS Medicine , v ol. 16, no. 1, Jan. 2019, doi: 10.1371/journal.pmed.1002730. [26] J. A. Golden, “Deep learning algorithms for detection of lymph node metastases from breast cancer: helping artificial intelligence be seen, ” J AMA , v ol. 318, no. 22, Dec. 2017, doi: 10.1001/jama.2017.14580. Deep learning for cate gorizing micr osatellite stability in color ectal cancer (Sofyan El Idrissi) Evaluation Warning : The document was created with Spire.PDF for Python.
3768 r ISSN: 2252-8938 [27] L. Zhang, L. Lu, I. Nogues, R. M. Summers, S. Liu, and J. Y ao, “DeepP ap: deep con v olutional netw orks for cervical cell classification, ” IEEE J ournal of Biomedical and Health Informatics , v ol. 21, no. 6, pp. 1633–1643, No v . 2017, doi: 10.1109/JBHI.2017.2705583. [28] B . K orbar et al. , “Deep learning for classification of colorectal polyps on whole-slide images, ” J ournal of P atholo gy Informatics , v ol. 8, no. 1, Jan. 2017, doi: 10.4103/jpi.jpi 34 17. [29] N. Coudray et al. , “Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning, ” Natur e Medicine , v ol. 24, no. 10, pp. 1559–1567, Oct. 2018, doi: 10.1038/s41591-018-0177-5. [30] J. N. Kather et al. , “Deep learning can predict microsatellite instability directly from histology in g astrointestinal cancer , ” Natur e Medicine , v ol. 25, no. 7, pp. 1054–1056, Jul. 2019, doi: 10.1038/s41591-019-0462-y . [31] M. Nalisnik et al. , “Interacti v e phenotyping of lar ge-scale histology imaging data with HistomicsML, ” Scientific Reports , v ol. 7, no. 1, No v . 2017, doi: 10.1038/s41598-017-15092-3. [32] P . M. K ulkarni et al. , “Deep learning based on standard H&E images of primary melanoma tumors identifies patients at risk for visceral recurrence and death, ” Clinical Cancer Resear c h , v ol. 26, no. 5, pp. 1126–1134, Mar . 2020, doi: 10.1158/1078-0432.CCR-19-1495. [33] J . K e, Y . Shen, Y . Guo, J. D. Wright, and X. Liang, “ A prediction model of microsatellite status from histology images, ” in 2020 10th International Confer ence on Biomedical Engineering and T ec hnolo gy (ICBET 2020) , 2020, pp. 334–338, doi: 10.1145/3397391.3397442. [34] R . Y amashita et al. , “Deep learning model for the prediction of micr osatellite instability in colorectal cancer: a diagnostic study , ” The Lancet Oncolo gy , v ol. 22, no. 1, pp. 132–141, Jan. 2021, doi: 10.1016/S1470-2045(20)30535-0. [35] M . Bilal et al. , “De v elopment and v alidation of a weakly supervised deep learning frame w ork to predict the status of molecular pathw ays and k e y mutations in colorectal cancer from routine histology images: a retrospecti v e study , ” The Lancet Digital Health , v ol. 3, no. 12, pp. e763–e772, Dec. 2021, doi: 10.1016/S2589-7500(21)00180-1. [36] K. Sirinukunw attana et al . , “Image-based consensus molecular subtype (imCMS) classification of colorectal cancer using deep learning, ” Gut , v ol. 70, no. 3, pp. 544–554, Mar . 2021, doi: 10.1136/gutjnl-2019-319866. [37] S. Elidrissi, I. B. A. Ouahab, M. Bouhorma, and F . Elouaai, “Un v eiling the clinical significance of microsatellite instability in colorectal cancer: deep learning and the se gment an ything model for accurate se gmentation and classification, ” International J ournal of Online and Biomedical Engineering , v ol. 21, no. 6, pp. 97–110, 2025, doi: 10.3991/ijoe.v21i06.54491. [38] J . Gibert, “TCGA CO AD MSI vs MSS prediction (JPG), ” Ka g gle . Accessed: Dec. 16, 2025. [Online]. A v ailable: https://www .kaggle.com/datasets/joangibert/tcg a coad msi mss jpg [39] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for im age recognition, ” in 2016 IEEE Confer ence on Computer V ision and P attern Reco gnition (CVPR) , Jun. 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90. [40] A. Ahmed, “Optimized ResNet-18 architecture for multi-class oral diseases classification, ” Karbala International J ournal of Modern Science , v ol. 12, no. 1, Jan. 2026, doi: 10.33640/2405-609X.3446. [41] S . B. Francis and J. P . V erma, “Deep CNN ResNet-18 based model with attention and transfer learning for Alzheimer’ s disease detection, ” F r ontier s in Neur oinformatics , v ol. 18, Jan. 2025, doi: 10.3389/fninf.2024.1507217. [42] Z. Du, X. Liu, J. Xu, G. Nan, and G. W ang, “Semantic reasoning with NLI for assertion detection in medical te xt, ” in 2022 IEEE International Confer ence on Bioinformatics and Biomedicine (BIBM) , Dec. 2022, pp. 744–748, doi: 10.1109/BIBM55620.2022.9995166. [43] H. Lee, J. Seo, G. Lee, J. P ark, D. Y eo, and A. Hong, “T w o-stage classification method for MSI status prediction based on deep learning approach, ” Applied Sciences , v ol. 11, no. 1, Dec. 2020, doi: 10.3390/app11010254. [44] J. K e, Y . Shen, J. D. Wright, N. Ji ng, X. Liang, and D. Shen, “Identifying patch-le v el MSI from histological images of colorectal cancer by a kno wledge distillation model, ” in 2020 IEEE International Confer ence on Bioinformatics and Biomedicine (BIBM) , Dec. 2020, pp. 1043–1046, doi: 10.1109/BIBM49941.2020.9313141. BIOGRAPHIES OF A UTHORS Pr of . Dr . Sofyan El Idrissi holds a doctor of Biotechnologies and Digital Healthcare de gree from Abdelmalek Essaadi Uni v ersity , Morocco in 2026. He is professor of biology at AREF Morocco, a researcher in oncology , a t the C3S Lab Computer Science and Smart Systems (C3S) Laboratory , F acul ty of Sciences and T echnologies of T angier , Uni v ersity Abdelmalek Essaadi, T angier , Morocco. Professor at “Institut Genomic Des Sciences De La Sante” professional training institute in the field of health and medical sciences. He can be contacted at email: idrissisofyan38@gmail.com. Dr . Y assine Drider is Doctor of Pharmac y , F aculty of Medicine and Pharmac y Rabat, Morocco. He is founder and o wner of pharmac y El Y ahiaouin. He is supervising professor at Institut Genomic Des Sciences De La Sante, institute of professional training in the field of health and medical sciences. He is doctoral researcher in oncology , at the Computer Science and Smart Systems (C3S) Laboratory , F aculty of Sciences and T echnologies of T angier , Uni v ersity Abdelmalek Essaadi, T angier , Morocco. He is an author of the book entitled: “La biologie quantique”. He can be contacted at email: yassinedrider@gmail.com. Int J Artif Intell, V ol. 15, No. 4, August 2026: 3761–3769 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 r 3769 Pr of . Dr . Ikram Ben Abdel Ouahab is an assistant professor in the Department of Computer Science at the F aculty of Sciences and T echniques of T angier , Morocco. She is a researcher and member of the data and intelligent systems (DIS) research team, at the Computer Science and Smart Systems (C3S) Laboratory , F aculty of Sciences and T echnologies of T angier , Uni v ersity Abdelmalek Essaadi, T angier , Morocco. She recei v ed her Ph.D. in Cybersecurity and AI from Abdelmalek Essaadi Uni v ersity . She has partici pated in numerous international conferences and has published se v eral scientific papers in collaboration with her laboratory colleagues. Her primary research interests include c ybersecurity , mal w are analysis, AI, AI in oncology , and the internet of things (IoT). She can be contacted at email: ibenabdelouahab@uae.ac.ma. Pr of . Dr . Mohammed Bouhorma is a full professor and e xperienced academic who has more than 25 years of teaching and tutoring e xperience in the areas of information security , security protocols, AI, big data, and digital forensics at F aculty of Sciences and T echnologies of T angier , Abdelmalek Essaadi Uni v ersi ty , T etouan, Morocco. He is member of research team at the Computer Science and Smart Systems (C3S) Laboratory , F aculty of Sciences and T echnologies of T angier , Uni v ersity Abdelmalek Essaadi, T angier , Morocco. His research interes ts include c yber -security , IoT , bigdata analytics, AI, smart cities technology , and serious g ames. He can be contacted at email: mbouhorma@uae.ac.ma. Pr of . Dr . F atiha El Ouaai is a full professor at F aculty of Sciences and T echnologies of T angier , Abdelmalek Essaadi Uni v ersity , T etouan, Morocco. Her research spans bioinformatics, human-computer interaction, computer communications, and computer security . She is member of research team a t the Computer Science and Smart Systems (C3S) Labora tory , F aculty of Sciences and T echnologies of T angier , Uni v ersity Abdelmalek Essaadi, T angier , Morocco. Her e xceptional contrib utions and publications ha v e earned her recognition, and she acti v el y participates in academic e v ents. Her research adv ances scientific kno wledge and dri v es inno v ation in multiple disciplines. She can be contacted at email: elouaaif@gmail.com. Deep learning for cate gorizing micr osatellite stability in color ectal cancer (Sofyan El Idrissi) Evaluation Warning : The document was created with Spire.PDF for Python.