Inter national J our nal of Inf ormatics and Communication T echnology (IJ-ICT) V ol. 15, No. 3, September 2026, pp. 1443 ∼ 1452 ISSN: 2252-8776, DOI: 10.11591/ijict.v15i3.pp1443-1452 ❒ 1443 A multi-cancer detection framew ork using deep lear ning and h ybrid machine lear ning appr oaches Karan Singh 1 , Amruta P awar 1 , Drish ya T omar 1 , Amrita Y ada v 1 , Aditi Chhabria 1 , V aibha v Narawade 2 1 Computer Science and Engineering, Ramrao Adik Institute of T echnology , D.Y . P atil Deemed to be Uni v ersity , Na vi Mumbai, India 2 Computer Engineering, Ramrao Adik Institute of T echnology , D.Y . P atil Deemed to be Uni v ersity , Na vi Mumbai, India Article Inf o Article history: Recei v ed Jul 7, 2025 Re vised Jun 15, 2026 Accepted Jul 5, 2026 K eyw ords: Hybrid classiers Medical imaging Multi-cancer detection ResNet50 T ransfer learning ABSTRA CT The diagnostic solutions of fered by the present articial intelligence (AI) solu- tions suf fer from non-generalizability and hea vy reliance on comple x models. In an attempt to solv e the se issues, we propose a lightweight yet v ersatile method consisting of a combination of ResNet50 transfer learning and h ybrid machine learning. Image features are e xtracted from dermoscopic, MRI, and histopatho- logical ima ges. These are subj ected to principal component analysis (PCA) di- mensionality reduction follo wed by classication using support v ector machine (SVM), random forest (RF), logistic re gression (LR), and XGBoost algorithms. This se gre g ation of the tw o processes impro v es ef cienc y . The h ybrid approach using ResNet50 + LR yielded an accurac y of 91.01% in the case of breast can- cer detection compared to 86.26% of a baseline con v olutional neural netw ork (CNN). Also, ResNet50 g a v e an accurac y of 96.61% in diagnosing skin cancer . Custom CNN pro vided an accurac y of 99.42% for lung cancer and 96.33% for brain tumor detection. This is an open access article under the CC BY -SA license . Corresponding A uthor: Karan Singh Computer Science and Engineering, Ramrao Adik Institute of T echnology D.Y . P atil Deemed to be Uni v ersity Nerul, Na vi Mumbai, India Email: karans9609@gmail.com 1. INTR ODUCTION Cancer causes close to 10 million deaths each year , which mak es it vital to ha v e f ast and reliable means of diagnosis [1]. Deep learning (DL) has automated cancer detection. Ho we v er , e xisting systems are often siloed, focusing on single cancer types, and require high computational res ources limiting clinical deplo yment. In this re g ard, we propose a h ybrid frame w ork that mer ges ResNet50 feature e xtraction with classical machine learning (ML). This e xploits the rob ustness of DL [2] and reduces computati o na l costs for scalable multi-cancer screening. V arious DL and h ybrid approaches for cancer detection ha v e been studied recently as summarized in T able 1. Nugroho et al. [3] used con v olutional neural netw ork (CNN) to classify skin lesion on HAM10000 b ut the y did not do proper preprocessing and achie v ed 78% accurac y . Nahata and Singh [4] achie v ed 91% accurac y using InceptionResNet on ISIC datasets, b ut f aced class imbalance. Ahmed et al. [5] used 3D CNN on LUN A16 dataset for lung cancer classication with an accurac y of 80%. GS-PCANet with principal component analysis (PCA), support v ector machine (SVM), Ram et al. [6] - 90.8%. Ramamoorth y et al. [7] used ResNet50/Xception in Moroccan datasets of breast cancer b ut the results were not generalized. J ournal homepage: http://ijict.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
1444 ❒ ISSN: 2252-8776 Asri et al. [8] used traditional ML models lik e KNN, SVM, and decision tree (DT) on WBCD dataset and achie v ed 86.9%. Cheng et al. [9] and Afs har et al. [10] used CNN and capsule netw orks respecti v ely for brain tumor classication with accurac y of 84-86%. Ho we v er , these promising approaches are mostly restricted to indi vidual cancer types and lack a unied multi-cancer frame w ork [11]. T able 1. Literature surv e y analysis for cancer detection Reference Cancer type Model(s) used Dataset Metrics K e y insights Limitations Nugroho et al. [3] Skin CNN HAM 10000 Accurac y: 78 CNN classies skin photos well No model comparison; fe w metrics; no preprocessing and v alidation. Nahata and Singh [4] Skin Inception ResNet ISIC 2018 + ISIC 2019 Accurac y: 91 High accurac y with DL, transfer learning, and augmentation. Class imbalance, less interpretable, lacks practical v alidation. Ahmed et al. [5] Lung V anilla 3D CNN LUN A16 Accurac y: 80 3D CNN preserv es spatial conte xt. Lo wer accurac y; fe w test slices; simple CNN used. Ram et al. [6] Lung GS- PCANet HE stained slides from K-rasG12D model Accurac y: 90.8 GS-PCANet with reduced PCA and SVM for lesion detection. F alse positi v es; noise-sensiti v e; computationally e xpensi v e. Ramamoorth y et al. [7] Breast ResNet50, Xception Pri v ate Moroccan Accurac y: 88 Xception better than ResNet50; both ha v e clinical promise. Sparse data, no v alidation; poor generalizability . Asri et al. [8] Breast KNN, SVM, DT W isconsin Breast Cancer (WBCD) Accurac y: 86.9 Basic ML algorithms; DT best. No DL; lo w accurac y; no h ybrid model. Cheng et al. [9] Brain CNN 3064 MRI images Accurac y: 84.19 Re gion augmentation impro v es CNNs on small datasets. Accurac y limited; comple x model. Afshar et al. [10] Brain Capsule Netw ork 3064 T1-weighted MRI images Accurac y: 86.56 Capsule netw orks handle spatial relationships better . Lo wer accurac y than CNNs; h yperparameter sensiti v e; compute intensi v e. Though DL has automated the process of diagnosis, current methods w ork independently with one cancer type at a time and ha v e high computational cost hindering their applicability in the clinic setting. W e propose a no v el methodology incorporating ResNet50 features e xtraction and traditional ML to e xploit the adv antages of DL technology with minimized computational e xpenses for multiple cancers detection. Research g ap: despite e xtensi v e research, e xisting studies generally lack: − A unied frame w ork for a common DL-ML pipeline for multiple cancer type detection. − Cross-modality generalization: testing the models on dermoscopic, MRI, and histopathological images. − A comparati v e h ybrid benchmark that systematically combines ResNet50 with classical ML classiers. T o address this, we propose Health-Buddy , a unied frame w ork that combines ResNet50 [12] for feature e xtraction, PCA dimensionality reduction, and h ybrid classiers (SVM [13], RF [14], LR [15], and XGBoost [16]), benchmark ed ag ainst a custom CNN [17]. T o conte xtualize the datasets and clinical importance, the follo wing cancer types are addressed: − Skin cancer: dermoscopic e xamination of dif ferent lesions. − Lung cancer: detection of malignant tumors in histopathology . − Breast cancer: microscopic dif ferentiation of benign and malignant tissue. − Brain tumor: MRI-based detection of glioma, pituitary , and meningioma. This di v ersity conrms the model’ s a bility to generalize across visually disparate data types, of fering a scalable, lightweight alternati v e to comple x end-to-end models. Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1443–1452 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1445 2. METHOD HealthBuddy emplo ys a consistent mult i-stage pipeline that includes data collection, pre-proces sing, custom CNNs, transfer learning with ResNet50, and h ybrid classication. 2.1. Data collection T able 2 lists the datasets utilized in this study , along with their description and size for each cancer type. These datasets of fer di v ersity across imaging modalities for the assessment of model generalizability . T able 2. Datasets used for each cancer type Cancer Dataset Description Size Skin HAM10000 [18] Dermatoscopic images classied into se v en lesion types 10,000 Lung LC25000 [19] Histopathological images of health y and cancerous tissues 15,000 Breast Breast Cancer -AE [20] Microscopic breast tissue images labeled benign / malignant 8,000 Brain Brain MRI [21] MRI scans of glioma, meningioma, pituitary , or no tumor 7,000 2.2. Data pr epr ocessing and augmentation The images were resized to 224 × 224 with 3 RGB channels. F or e v aluation, to ha v e equal number of images from each class, stratied sampling w as performed to di vide the dataset into 70%, 15%, and 15% train, v alidate and test respecti v ely . Data augmentation of training images included scaling ( 1 / 255 ), ± 20 ◦ rotation, horizontal ipping and shear zooming ( ≤ 10% ) by nearest neighbor interpolation. No dat a augmentation w as performed on v alidation and test images. 2.3. Custom CNN ar chitectur e A custom CNN model w as de v eloped to perform multiclass classication of multiple cancer types . The proposed architecture, illustrated in Figure 1, is designed to automatically e xtract hierarchical features from input images and classify them into their respecti v e cancer cate gories. The architecture consists of three con v olutional blocks for feature e xtraction, follo wed by a atten layer , a fully connected layer with dropout re gularization, and a SoftMax output layer for nal classication. Figure 1. Custom CNN architecture used for all cancer types 2.3.1. Con v olutional blocks (featur e extraction) Three Con v2D layers ( 3 × 3 k ernels with 32, 64, 128 lters) were used to capture hierarchical features. Each con v olution is immediately follo wed by: − ReLU acti v ation: introduces non-linear functions to model intricate relationships. ReLU(x) = max(0, x) − Batch normalization: stabilizes the learning process and helps it con v er ge f aster . − Max pooling ( 2 × 2 ): do wnsamples the input feature maps to decrease computations. A multi-cancer detection fr ame work using deep learning and hybrid mac hine learning ... (Kar an Singh) Evaluation Warning : The document was created with Spire.PDF for Python.
1446 ❒ ISSN: 2252-8776 2.3.2. Classication head − Flattening: once features ha v e been e xtracted, the 2D feature maps are attened to become 1D. − Fully connected layer: the 1D v ector is passed through a fully connected layer (128 units) with applying L2 re gularization ( λ = 0 . 001 ) and dropout (rate=0.5) to pre v ent o v ertti ng. − Classication: the output layer emplo ys SoftMax function acti v ation to pro vide class probabilities specic to cancer types (e.g., 7 classes for skin cancer , 4 for brain cancer). SoftMax ( z i ) = e z i P K j =1 e z j , for i = 1 , 2 , . . . , K (1) 2.3.3. T raining conguration − Optimizer: Adam with a learning rate of 1 × 10 − 4 − Loss function: cate gorical crossentrop y − Epochs: trained for up to 50 epochs − Callbacks: to impro v e the con v er gence, early stopping (with patience = 10) and learning rate reduction (f actor = 0.5, patience = 5) were used. 2.4. T ransfer lear ning with ResNet50 ResNet50 pre-trained on ImageNet w as used taking adv antage of its 50 residual layers to get rob ust hierarchical fea tures. Figure 2 sho ws the ResNet50 modied architecture used for joint detection of all four cancer classes. Figure 2. T ransfer learning architecture: modied v ersion of ResNet50 ha ving frozen layers, ne tuned layers, and a customized classication head 2.4.1. F eatur e extraction (fr ozen base) Initially , con v olutional layers were frozen to preserv e lo w and mid le v el features. After that, a cus- tom classication head w as created, which consisted of GlobalA v eragePooling2D, Dense (256 units, ReLU), BatchNormalization and Dropout ( r ate = 0 . 5 ) to reduce o v ertting. The classication head w as trained for 30 epochs. 2.4.2. Fine-tuning phase Then the last 20 layers were ne-tuned for 20 epochs using Adam optimization (learning rate = 1 × 10 − 5 ). F or better con v er gence and dealing with class imbalance, class weights, Earl y S topping (patience=10), and ModelCheckpoint were used. 2.4.3. Hybrid model justication The selected features are reduced using PCA [22] before being classied with SVM, RF , LR, or XG- Boost. Compared to computationally intensi v e visi on transformers (V iTs), this lightweight approach reduces training comple xity and memory usage during inference. Pre vious studies ha v e sho wn that such h ybrid models can outperform standalone CNNs in li v er disease detection [23] and skin cancer classication [24]. Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1443–1452 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1447 2.5. Deployment via ask T o e v aluate its practical applicability , the models were deplo yed through a ask-based web application [25], enabling real-time image analysis and interacti v e clinical testing [26]. 3. RESUL TS AND DISCUSSION W e e v aluated the custom CNN, ResNet50, and h ybrid architectures (SVM, RF , LR, and XGBoost) across four cancer types using standard metrics: accurac y , precision, recall, and F1-score. 3.1. Accuracy The comparison is sho wn in T able 3. Although the ResNet50 model w as the best performer on skin cancer dat aset (96.61%), h ybridization of ResNet50 and LR achie v ed better results than just CNN (86.26%) in breast cancer . T able 3. Accurac y of v arious models on dif ferent cancers Model Skin Lung Breast Brain CNN 93.22% 99.42% 86.26% 96.33% ResNet 96.61% 92.31% 86.26% 93.44% ResNet + SVM 93.89% 91.24% 88.63% 93.82% ResNet + RF 95.06% 91.60% 85.68% 93.82% ResNet + LR 95.50% 91.24% 91.01% 94.36% ResNet + XGB 93.94% 91.96% 90.04% 93.75% 3.2. Pr ecision T able 4 demonstrates the capability of the proposed model to reduce the number of f alse posit i v es. The h ybri d models that used LR and XGBoost achie v ed precision scores of ( ≥ 91% ) on the lung, breast, and brain datasets. T able 4. Precision of v arious models on dif ferent cancers Model Skin Lung Breast Brain CNN 94% 99% 84% 94% ResNet 97% 92% 84% 93% ResNet + SVM 95% 91% 92% 94% ResNet + RF 96% 92% 85% 94% ResNet + LR 96% 91% 94% 94% ResNet + XGB 94% 92% 91% 94% 3.3. Recall T able 5 demonstrates the sensit i vity to w ards positi v e cases. The custom CNN had almost perfect recalls on both lung and breast cancer (99%), whereas the h ybrid model s presented better balance of sensiti vity and specicity . T able 5. Recall of v arious models on dif ferent cancers Model Skin Lung Breast Brain CNN 94% 99% 99% 92% ResNet 97% 92% 99% 93% ResNet + SVM 94% 91% 91% 93% ResNet + RF 95% 91% 98% 93% ResNet + LR 96% 91% 93% 94% ResNet + XGB 94% 91% 95% 93% 3.4. F1-scor e As can be seen from T able 6, h ybrid approaches retain good precision-recall balance re g ardless of dataset imbalance. This observ ation supports the reliability of the feature e xtraction process based on ResNet50 netw ork [27]. A multi-cancer detection fr ame work using deep learning and hybrid mac hine learning ... (Kar an Singh) Evaluation Warning : The document was created with Spire.PDF for Python.
1448 ❒ ISSN: 2252-8776 T able 6. F1-score of v arious models on dif ferent cancers Model Skin Lung Breast Brain CNN 92% 99% 91% 94% ResNet 95% 92% 91% 93% ResNet + SVM 95% 91% 92% 93% ResNet + RF 95% 91% 90% 93% ResNet + LR 95% 91% 93% 94% ResNet + XGB 95% 91% 93% 93% 3.5. T raining and v alidation accuracy and loss cur v es Figures 3-6 illustrate the learning curv es for the best-performing models. Across all cancer types, the graphs sho w consistently rising accurac y and decreasing loss, indicating ef fecti v e con v er gence. The lung cancer CNN Figure 4, in particular , demonstrates stable learning with minimal o v ertting, v alidating the ef cac y of the chosen h yperparameters. Figure 3. Skin cancer (ResNet50) Figure 4. Lung cancer (CNN) Figure 5. Breast cancer (ResNet50+LR) Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1443–1452 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1449 Figure 6. Brain cancer (CNN) 3.6. Confusion matrix analysis Figure 7 conrms high classication reli ability . The brain CNN Figure 7(a) achie v es high sensiti vity for meningioma, despite minor o v erlap between glioma and pituitary classes. Breast Figure 7(b) and lung Figure 7(c) models demonstrate clear separation between malignant and benign/health y tissues. Skin ResNet50 Figure 7(d) e xhibits a strong diagonal trend, ef fecti v ely isolating melanoma from benign lesions. (a) (b) (c) (d) Figure 7. Confusion matrices of the best performing model for each cancer type; (a) CNN for brain, (b) ResNet50+LR for breast, (c) CNN for lung, and (d) ResNet50 for skin 3.7. Comparati v e analysis and discussion Figure 8 conrms that optimal architecture relies on the imaging modality . Consistent with literat ure [10], [14], our custom CNN e xcelled in high-contrast lung (99.42%) and brain (96.33%) tasks . Con v ersely , the h ybrid ResNet50+LR achie v ed superior performance for breast histopathology (91.01%) [11], [21]. A multi-cancer detection fr ame work using deep learning and hybrid mac hine learning ... (Kar an Singh) Evaluation Warning : The document was created with Spire.PDF for Python.
1450 ❒ ISSN: 2252-8776 Error analysis re v eals specic challenges, such as MRI intensity o v erlaps between Glioma and Pituitary tu- mors. Ho we v er , compared to resource-intensi v e V iTs, our h ybrid approach of fers a strate gic trade-of f. It achie v es competiti v e accurac y (e.g., 96.61% for Skin) with a signicantly lighter computational footprint, v al- idating its suitability for rapid, resource-constrained screening. Figure 8. Comparati v e accurac y of deep learning and h ybrid models across dif ferent cancers 3.8. Ethical and clinical deployment challenges The clinical application needs to deal with the demographic bias in the dataset (for e xample, HAM10000). In order to solv e the problem related to the “black box” system, we used an interpretable classier lik e logistic re gression in our h ybri d layer . In addition, our light model allo ws inference from de vice- le v el Edge AI. 4. CONCLUSION The current study creates a unied frame w ork for multi-cancer screening, demonstrating the ef fecti v e- ness of the same h ybrid model (ResNet50 + ML) in dif ferent imaging techniques. It is clear from our results that although CNN w orks well when there is a need for high contrast (lung/brain), h ybrid approaches outper - form the former when it comes to dealing with compl icated te xtures (breast). Combining high accurac y (up to 96.61%) with reduced costs, the frame w ork becomes applicable in clinical practice. FUNDING INFORMA TION Authors state no funding in v olv ed. 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 Karan Singh ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Amruta P a w ar ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Drish ya T omar ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Amrita Y ada v ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Aditi Chhabria ✓ ✓ ✓ V aibha v Nara w ade ✓ 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 Administrati on 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 Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1443–1452 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1451 CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The datasets supporting the ndings of this study are publicly a v ailable and ha v e been referenced appropriately within the manuscript. Specically , four open-access datasets were used for skin, breast, brain, and lung cancer classication tasks. All dataset sources are cited in the References section. REFERENCES [1] W orld Health Or g anization, “Cancer , ” W orld Health Or g anization. Accessed: Jul. 20, 1BC. [Online]. A v ailable: https://www .who.int/ [2] G. Litjens et al. , “ A surv e y on deep learning in medical image analysis, ” Medical Ima g e Analysis , v ol. 42, no. 103994, pp. 60–88, Dec. 2017, doi: 10.1016/j.media.2017.07.005. [3] A. A. Nugroho, I. Slamet, and Sugiyanto, “Skins cancer identication system of HAMl0000 skin cancer dataset using con v olutional neural netw ork, ” in AIP Confer ence Pr oceedings , 2019, p. 020039. doi: 10.1063/1.5141652. [4] H. Nahata and S. P . Singh, “Deep learning solutions for skin cancer detection and diagnosis, ” in Learning and Analytics in Intellig ent Systems , v ol. 13, 2020, pp. 159–182. doi: 10.1007/978-3-030-40850-3 8. [5] T . Ahmed, M. S. P arvin, M. R. Haque, and M. S. Uddin, “Lung cancer detection using CT image based on 3D con v olutional neural netw ork, ” J ournal of Computer and Communications , v ol. 08, no. 03, pp. 35–42, 2020, doi: 10.4236/jcc.2020.83004. [6] S. Ram et al. , “Lung cancer lesion detection in histopathology images using graph-based sparse PCA netw ork, ” Neoplasia (United States) , v ol. 42, p. 100911, Aug. 2023, doi: 10.1016/j.neo.2023.100911. [7] P . Ramamoorth y , B. R. R. Reddy , S. S. Askar , and M. Abouha ww ash, “Histopathology-based breast cancer pre diction using deep learning methods for healthcare applications, ” F r ontier s in Oncolo gy , v ol. 14, Jun. 2024, doi: 10.3389/fonc.2024.1300997. [8] H. Asri, H. Mousannif, H. Al Moatassime, and T . Noel, “Using machine learning algorithms for breast cancer risk prediction and diagnosis, ” Pr ocedia Computer Science , v ol. 83, pp. 1064–1069, 2016, doi: 10.1016/j.procs.2016.04.224. [9] J. Cheng et al. , “Enhanced performance of brain tumor classication via tumor re gion augmentation and partition, ” PLOS ONE , v ol. 10, no. 10, p. e0140381, Oct. 2015, doi: 10.1371/journal.pone.0140381. [10] P . Afshar , A. Mohammadi, and K. N. Plataniotis, “Brain tumor type classication via capsule netw orks, ” Pr oceedings - International Confer ence on Ima g e Pr ocessing , ICIP , pp. 3129–3133, 2018, doi: 10.1109/ICIP .2018.8451379. [11] H. M. Rai, “Cancer detection and se gmentation using machine learning and deep learning techniques: a re vie w , ” Multimedia T ools and Applications , v ol. 83, no. 9, pp. 27001–27035, 2024, doi: 10.1007/s11042-023-16520-5. [12] K. He, X. Zhang, S. Ren, and J. Sun, “De ep residual learning for image recognition, ” Pr oceedings of the IEEE Computer Society Confer ence on Computer V ision and P attern Reco gnition , v ol. 2016-December , pp. 770–778, 2016, doi: 10.1109/CVPR.2016.90. [13] C. Cortes and V . V apnik, “Support-v ector netw orks, ” Mac hine Learning , v ol. 20, no. 3, pp. 273–297, Sep. 1995, doi: 10.1007/BF00994018. [14] L. Brei man, “Random forests, ” Mac hine Learning , v ol. 45, no. 1, pp. 5–32, Oct. 2001, doi: 10.1023/A:1010933404324. [15] A. Stuart and D. R. Cox, Analysis of Binary Data , v ol. 38, no. 150. London, U.K.: Chapman and Hall, 1971. doi: 10.2307/2552586. [16] T . Chen and C. Guestrin, “XGBoost: a scalable tree boosting system, ” in Pr oceedings of the A CM SIGKDD International Confer ence on Knowledg e Disco very and Data Mining , Ne w Y ork, NY , USA: A CM, Aug. 2016, pp. 785–794. doi: 10.1145/2939672.2939785. [17] Y . LeCun, L. Bottou, Y . Bengio, and P . Haf fner , “Gradient-based learning applied to document recognition, ” Pr oceedings of the IEEE , v ol. 86, no. 11, pp. 2278–2323, 1998, doi: 10.1109/5.726791. [18] P . Tschandl et al. , “The HAM10000 dataset, a lar ge collection of multi-source dermatoscopic images of common pigmented skin lesions, ” Natur e Medicine , v ol. 26, no. 8, pp. 1229–1234, 2020. [19] A. A. Bork o wski, M. M. Bui, L. B. Thomas, C. P . W ilson, L. A. DeLand, and S. M. Mastorides, “Lung and colon cancer histopatho- logical image dataset (LC25000), ” arXiv pr eprint arXiv:1912.12142 , 2019. [20] F . A. Spanhol, L. S. Oli v eira, C. Petitjean, and L. Heutte, “ A dataset for breast cancer histopathological image classication, ” IEEE T r ansactions on Biomedical Engineering , v ol. 63, no. 7, pp. 1455–1462, 2016, doi: 10.1109/TBME.2015.2496264. [21] M. Nickparv ar , “Brain tumor MRI dataset, ” in Ka g gle , 2020. [Online]. A v ailable: https://www .kaggle.com/datasets/masoudnickparv ar/brain-tumor -mri-dataset [22] A. G eron, Hands-On Mac hine Learning with Scikit-Learn, K er as, and T ensorFlow: Concepts, T ools, and T ec hniques to Build Intellig ent Systems , 2nd ed. Se. CA, USA: O’Reilly Media, 2019. [23] Pe ter and Ris, “Li v er cancer classication via deep h ybrid model from CT images using CNN and SVM, ” 2025, Resear c hGate . [24] R. Maurya, A. K. Bais, T . Gopalakrishnan, M. K. Dutta, N. N. P ande y , and Y . V . S. Murth y , “Skin lesion classication using deep feature fusion and selection using XGBoost classier , ” in 2024 IEEE International Students’ Confer ence on Electr ical, Electr onics and Computer Science , SCEECS 2024 , IEEE, Feb . 2024, pp. 1–5. doi: 10.1109/SCEECS61402.2024.10481955. [25] M . Grinber g, Flask W eb De velopment: De veloping W eb Applications with Python , 2nd ed. Se. CA: O’Reilly Media, 2018. [26] S. Narayanan, N. M. Balamurug an, K. Maithili, and P . B. P alas, “Le v eraging machine learning methods for multiple disease pre- diction using p ython ML libraries and ask API, ” in Pr oceedings - International Confer ence on Applied Articial Intellig ence and Computing , ICAAIC 2022 , IEEE, May 2022, pp. 694–701. doi: 10.1109/ICAAIC53929.2022.9792807. [27] R . Y acouby and D. Axman, “Probabilistic e xtension of precision, recall, and F1 score for more thorough e v aluation of classication models, ” in Pr oceedings of the F ir st W orkshop on Evaluation and Comparison of NL P Systems , Stroudsb ur g, P A, USA: Association for Computational Linguistics, 2020, pp. 79–91. doi: 10.18653/v1/2020.e v al4nlp-1.9. A multi-cancer detection fr ame work using deep learning and hybrid mac hine learning ... (Kar an Singh) Evaluation Warning : The document was created with Spire.PDF for Python.
1452 ❒ ISSN: 2252-8776 BIOGRAPHIES OF A UTHORS Karan Singh is a nal-year bachelor of T echnology student specializing in Articial Intelligence and Data Science at Ramrao Adik Institute of T echnology , Na vi Mumbai. His research focuses on machine learning and deep learning. Professionally , he serv es as a Softw are De v elopment Engineer (SDE) at V ector Consult ing Group, ha ving pre viously completed an SDE internship there. He can be contacted via email at karans9609@gmail.com. Amruta P awar is pursuing bachelor of technology in Articial Intelligence and Data Science at Ramrao Adik Ins titute of T echnology . Her research interests are in machine learning and AI-dri v en healthcare applications. She has completed an internship at L TIMindtree focusing on asset management and data analysis. Her interests include articial intelligence, machine learning, computer vision, and data science. She can be contacted at email: amrutapa w ar004@gmail.com. Drish ya T omar is pursuing bachelor of technology in Articial Intelligence a nd Data Science at Ramrao Adik Institute of T echnology . She has completed internships as an IT Solutions Intern at V irtusa and Machine Learning Intern at IDEALERS B2B PVT L TD. Her academi c interests include machine learning, deep learning ,computer vision data intelligence, and ethical hacking. She can be contacted at email: drish yatomar2024@gmail.com. Amrita Y ada v is current ly pursuing a bachelor of technology in Articial Intelligence and Data Science at Ramrao Adik Institute of T echnology . Her w ork focuses on deep learning and machine learning approaches. She is currently an intern at Deloitte, where she w orks on Salesforce and service-related solutions. She has also completed her tenure with the National Cadet Corps. She can be contacted at email: amrita22205@gmail.com. Aditi Chhabria is currently an associate profess or in the Department of Computer Science and Engineering at Ramrao Adik Institute of T echnology , Dr . D. Y . P atil Deemed to be Uni v ersity , Na vi Mumbai, India. W ith o v er 20 years of teaching e xperience, Dr . Chhabria specializes in data science and machine learning. She has co-authored se v eral research papers in these domai ns. She can be contacted at email: aditi.chhabria@rait.ac.in. V aibha v Narawade is a professor at Ramrao Adik Institute of T echnology , Nerul. He has o v er 24 years of teaching and administrati v e e xperience. He has serv ed as a Membe r of the Senate (2008-2010) and as a Member of the Management Council and Senate (2017-2022) at the Uni v ersity of Mumbai. He is also a member of the Board of Studies in CSE, a Member of the Re- search Recognition Committee at Pun yashlok Ahilyade vi Holkar Uni v ersity , and a F aculty Member at Dr . Babasaheb Ambedkar T echnological Uni v ersity . He has authored or co-authored se v eral in- ternational journal and conference papers, book chapters, and four books. He re gularly serv es as a session chair and re vie wer for international conferences and is on the editori al board of the American Journal of Agricultural Science, Engineering, and T echnology (AJ ASET). He is also a judge for the A vishkar Research Con v ention and is a fello w of ISTE and A CM. He can be contacted a t email: v aibha v .nara w ade@rait.ac.in. Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1443–1452 Evaluation Warning : The document was created with Spire.PDF for Python.