IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 14, No. 1, February 2025, pp. 54 ∼ 61 ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i1.pp54-61 ❒ 54 Pr edicting enhanced diagnostic models: deep lear ning f or multi-label r etinal disease classication Sridhe vi Sundararajan, Harikrishnan Ramachandran, Harshita Gupta, Y ashraj P atil Symbiosis Institute of T echnology (SIT)-Pune Campus, Symbiosis International Deemed Uni v ersity (SIDU), Pune, India Article Inf o Article history: Recei v ed Mar 29, 2024 Re vised Jun 28, 2024 Accepted Jul 26, 2024 K eyw ords: Classication Deep learning Medical imaging Multi-label Prediction models Retinal disease ABSTRA CT In this study , we assess three con v olutional neural netw ork (CNN) architec- tures—V GG16, ResNet50, and InceptionV3 for multi class ication of fundus images in the retinal fundus multi-disease image dataset (RFMID2), compris- ing of 860 images. F ocusing on diabetic retinopath y , e xudation, and hem- orrhagic retinopath y , we preprocessed the dataset for uniformity and balance. Using transfer learning, the models were adapted for feature e xtraction and ne-tuned to our multi-label classicat ion task. Their performance w as mea- sured by subset accurac y , precision, recall, F1-score, hamming loss, and Jaccard score. V GG16 emer ged as the top performer , with the highest subset accurac y (84.81%) and macro precision (95.83%), indicating its superior class distinc- tion capabilities. Res Net50 sho wed commendable accurac y (79.75%) and preci- sion (86.70%), whereas InceptionV3 lagged with lo wer accurac y (66.67%) and precision (81.21%). These ndings suggest V GG16’ s depth of fers adv antages in multi-label classication, highlighting InceptionV3’ s limitations in comple x scenarios. This analysis helps optimize CNN architecture selection for specic tasks, sugges ting future e xploration of dataset v ariability , ensemble methods, and h ybrid models for impro v ed performance. This is an open access article under the CC BY -SA license . Corresponding A uthor: Harikrishnan Ramachandran Symbiosis Institute of T echnology (SIT)-Pune Campus, Symbiosis International Deemed Uni v ersity (SIDU) Pune 412115, India Email: harikrishnan.r@sitpune.edu.in 1. INTR ODUCTION The inte gration of AI-dri v en image processing in healthcare has mark ed a signicant leap forw ard in diagnosing critical retinal conditions such as diabetic ret inopath y [1], e xudation [2], and hemorrhagic retinopa- th y [3]. These conditions, notorious for their potent ial to impair vision, are increasingly being detected with adv anced computational approaches, notably con v olutional neural netw orks (CNNs) [4]. Our in v estig ation e v aluates the capabilities of three distinguished CNN models— V GG16 [5] , ResNet50 [6], and InceptionV3 [7]—on the retinal fundus multi-disease image dataset (RFMID2) dataset [8], which encompasses 860 fundus images annotated for these specic diseases. The generation of the dataset RFMID2 is depicted in Figure 1. Applying transfer learning [9] follo wed by feature e xtraction [10] of the fundal images, we aim to ne-tune these models to the task of multi-label retinal disease classication. The RFMID2 dataset serv es as an ideal ground for thi s e xploration, gi v en its di v erse and comprehensi v e annotation of retinal pathologies. By ensuring a balanced dataset and emplo ying rigorous preprocessing measures, we seek to establish a le v el eld for e v al- uating model performance across k e y metrics [11]: s ubset accurac y , precision, recall, F1 score, hamming loss, and Jaccard score [12]. As the global incidence of retinal diseases escalates, the ur genc y for de v eloping rapid J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 55 and reliable diagnosti c tools becomes paramount. This study not only compares the ef fecti v eness of dif ferent CNN architectures in automated retinal disease detecti on b ut also aims to contrib ute to the enhancement of diagnostic processes in ophthalmology . Through our ndings, we anticipate setting the stage for future ad- v ancements in automated retinal imaging [13], potenti ally transforming patient care and management in the eld. Figure 1. Creation of retinal fundus multi image dataset RFMID2 [8] The adv ent of CNNs has signicantly impacted machine learning [14], enhancing multi-label clas - sication [15] tasks across v arious domains, including medical image analysis [16]. V GG16, ResNet, and Inception are notable architectures that ha v e contrib uted to adv ancements in this eld. V GG16 is celebrated for its depth and the use of small, uniform lters, making it adept at capturing detailed features crucial for medical imaging. Its architectural simplicity has played a pi v otal role in deepening our understanding of CNNs. ResNet introduces skip connections, allo wing the training of v ery deep netw orks by o v ercoming the v anishing gradi- ent problem. This inno v ation enables the learning of comple x patterns, illustrating the model’ s v ersatility in tasks such as disease classication from medical scans. Inception models, with their unique inception modules, capture information at dif ferent scales. This capability is particularly benecial for medical diagnostics, where images may contain intricate patterns across v arious resolutions. Applying these CNNs in m edical imaging has transformed diagnostic methods, enabling the classication and diagnosis of multiple diseases from a sin- gle image with unprecedented accurac y . This adv ancement not only simplies the diagnostic process b ut also impro v es early disease detection, potentially enhancing patient outcomes. The continuous e v olution of CNN architectures promises further adv ancements in medical image analysis. V GG16, ResNet, and Inception mod- els, with their v ersatility and adaptability , promise a future where AI-dri v en diagnostics combined with human e xpertise, of fering f aster and more precise diagnoses. T able 1 describes the a brief re vie w on the deep learning techniques for retinal diseases classication. In the e xploration of CNNs for retinal disease cl assication using datasets lik e RFMID2, researchers f ace a set of challenges b ut also encounter numerous opportunities for adv ancement. Understanding these can guide future research directions and help in o v ercoming e xisting limitations. The T able 2 e xplains the challenges and opportunities in retinal disease classication right from data collection till diagnosis. The adv ent of deep le arning has re v olutionized the eld of medical imaging, of fering un pr ecedented opportunities for enhancing diagnostic accurac y and patient care. CNNs, in particular , ha v e sho wn remark- able success in automatically detecting and classifying diseases from v arious types of medical images. This potential is especially pertinent in the realm of ophthalmology , where conditions lik e diabetic retinopath y , e x- udation, and hemorrhagic retinopath y can be identi ed through retinal imaging. T imely and precise detection of these diseases is vital to pre v ent se v ere consequences, such as blindness. Thus, le v eraging deep learning for this purpose addresses a signicant need in healthcare. In light of these de v elopments, this w ork proposes to apply and compare three stat e-of-the-art CNN architectures—V GG16, InceptionV3, and ResNet 50—for the classication of ophthalmic diseases in retinal images. By le v eraging these models’ pre-trained weights for feature e xtraction and emplo ying custom classication layers, this study aims to in v estig ate each architecture’ s Pr edicting enhanced dia gnostic models: deep learning for multi-label ... (Sridhe vi Sundar ar ajan) Evaluation Warning : The document was created with Spire.PDF for Python.
56 ❒ ISSN: 2252-8938 ef cac y in accurately identifying diabetic retinopath y , e xudation, and hemorrhagic retinopath y . This compar - ati v e analysis will contrib ute to identifying the most suitable deep learning approach for enhancing diagnostic processes in ophthal mology , ultimately aiding in early disease detection and treatment planning. The T able 3 e xplains the techniques and analysis done for this research. T able 1. Re vie w on deep learning approaches in retinal disease classication T opic Description Multile v el glo ww orm sw arm optimization con v olutional neural netw ork (MGSCNN) [17] for multi-disease classication A method that inte grates CNN with glo ww orm sw arm optimization (GSO) [18] for classier optimization, emphasizing normalization, smoothing, and resizing in preprocessing. This model sho wcases the ef cac y in enhancing classication accurac y for retinal images. Addressing class imbalance [19] The literature pro vides v arious strate gies to handle class imbalance prob- lems. These strate gies include resampling methods (both o v ersampling minority classes and undersampling majority ones), classier adaptation that requires the model to be tailored to address the dataset’ s imbalance, ensemble approaches that use multiple models, and cost-sensiti v e methods that emplo y custom met- rics in the loss function to emphasize the minority classes. Optical coherence tomograph y (OCT) image classication with V GG-16 [20] Describes the adaptation of V GG-16 using transfer learning for OCT retinal im- ages, focusing on data augmentation and custom layers for disease classication. Includes performance e v aluation metrics and Grad-CAM for model visualization [21]. Publicly a v ailable fundus image datasets [22] A v ariety of datasets ha v e been re vie wed and are used in multi-label retinal disease classication, including DRIVE, ST ARE, CHASE-DB, Messidor , and e-ophtha data sets. Each dataset has been de v eloped for dif ferent diagnostic pur - poses, e xhibiting di v erse v ariations and abnormalities that are crucial for training deep learning models. T able 2. From data to diagnosis: articial intelligence challenges and opportunities in retinal disease classication Challenges Opportunities Class imbalance: A common issue in medical datasets, includ- ing those for ret inal diseases, is class imbalance. Minority classes (rarer diseases) are not much in number , which can bias the model to w ards more frequent conditi ons. Addressing this requires inno v ati v e strate gies be yond simple resampling, demanding more sophisticated approaches lik e cost-sensiti v e learning or synthetic data generation. Adv anced archi tectural inno v ations [23]: Exploring be yond V GG16, ResNet, and Inception models to ne wer architectures can pro vide opportunities for impro v ed performance. Inno v a- tions in neural netw ork desi gn could of fer solutions to class imbalance, enhance feature e xtraction, and impro v e model in- terpretability . Data quality and v ariability: The quality of retinal images can v ary signicantly due to dif ferent imaging conditions, patient demographics, and disease stages. This v ariability can hinder the model’ s ability to generali ze from the training data to real- w orld applications. Multi-modal data fusion [24]: Combining retinal fundus im- ages with other types of medica l data (e.g., OCT [25], pa- tient demographics) through mult i-modal learning approaches could enhance diagnostic accurac y and of fer a more holistic vie w of patient health. Interpretability and trust: F or clinical applications, it’ s crucial that model predictions are interpretable by healthcare profes- sionals. De v eloping models that of fer transparenc y in their decision-making process remains a challenge, af fecting their adoption in clinical settings. T ransfer learning and domain adaptation [26]: Le v eraging transfer learning more ef fecti v ely , including domain adaptation techniques, can help models better generalize across dif ferent datasets and imaging conditions, reducing the g ap between re- search settings and real-w orld applications. Inte gration with clinical w orko ws: Successfully inte grating AI models into clinical w orko ws in v olv es challenges related to system compatibility , data pri v ac y , and the need for real- time analysis without disrupting e xisting procedures. Collaboration between AI researchers and clinicians: Strength- ening the collaborat ion between AI researchers and clinical practitioners can lead to the de v elopment of more practical, user -friendly AI tools that t seamlessly into clinical w ork- o ws and address real healthcare needs. The study e xplores the impact of architectural v ariations inherent to V GG16, inceptionV3, and ResNet- 50. This includes the depth of the netw orks, the use of inception modules v ersus residual connections, and their implications for feature representation and e xtraction capabilities. Each model w as ne-tuned to the specic task of medical image classication. This in v olv ed adjusting learning rates, e xperimenting with dif ferent num- bers of dense layers in the classication head, and e xploring dropout for re gularization to pre v ent o v ertting. Int J Artif Intell, V ol. 14, No. 1, February 2025: 54–61 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 57 T able 3. CNN architectures in ophthalmic disease detection: techniques, v ariations, and insights Methodology/feature e xtraction Description Dataset description The study uses a RFMID2 consisting of 860 images labeled with the presence or absence of diabetic retinopath y , edema, and hemorrhagic retinopath y . The l abels are encoded in a CSV le, f acilitating a structured approach to supervised learning. Balancing and sampling T o address potential class imbalance, a bala nced subset of images w as constructed. This ensures equi- table representation of each condition, enhancing the model’ s ability to learn from an e v enly distrib uted dataset. Image resizing and nor - malization All images were resized to 512x512 pix els to standardize input size. Additionally , pix el v alues were normalized to a [0, 1] range to aid in model con v er gence and ef cienc y during training. Data analysis The comparati v e analysis of label frequencies and combinations w as conducted for both retinal fundus multi image datasets that includes creating graphical representations to illustrate label frequencies and co-occurrence patterns Utilization of pre-trained models Le v eraging the V GG16, InceptionV3, and ResNet50 models pre-trained on ImageNet enables the e x- traction of rich, hierarchical features from the medical images, e v en with a limited dataset size. Global a v erage pooling (GAP) F ollo wing feature e xtraction, a GAP layer w as applied to reduce the feature dimensionality while main- taining spatial hierarchies, f acilitating a rob ust input for the classication layer . 2. METHOD This section thoroughly outlines the methods adopted in our in v estig ation. It details the r esearch frame w ork, the procedures for obtaining the dataset, and the anal ytical strate gies applied to fulll our in v es- tig ati v e goals. W e of fer an in-depth description of our procedural approach to maintain openness and ensure a clear understanding of our research e x ecution. 2.1. Setup and v ariables consider ed The T able 4 describes the specication of the retinal fundus dataset used in the research. The RFMID2 [8] a publicly accessi ble dataset has 860 retinal images, respecti v ely . From this diabetic retinopath y , e xudation, and hemorrhagic retinopath y ha v e been identied for this reserach. T able 4. Specication of the retinas fundus dataset [8] Aspect Desciption of the medical data Research domain Multiple label classication of retinal fundus image Detailed study focus Classication of v arious diseases using retinal fundus images Collection method Utilized the T OPCON TRC-NW300 for data collection Data presentation Includes images and associated data in CSV format V ariables under study The majority of the subjects underwent mydriasis dilation using 0.5% tropicamide. Another subsect of patients without mydriasis were also accessed. Experimental setup Images captured using non-in v asi v e techniques, maintaining specic distances between the camera lenses and the subjects’ e yes with the patients in a seated position. Data origin Data collected from Shri Ganpati Netralaya in Jalna, Maharashtra, and the Center of Excellence in Signal and Image Processing at SGGS Institute of Engineering and T echnology in Nanded, Maharashtra, India. 2.2. Dataset pr ocessing The Figure 2 depicts the o wchart e xplaining the methodology used in this research for the classi ca- tion of retinal diseases from RFMID2 dataset. The methodology depicted in the o wchart e xplains the use of con v olutional neutal netw orks for the classication of retinal images. It starts with identifying the labels that are crucial for this study nam ely diabetic retinopath y , e xudation, and hemorrhagic retinopath y from the 860 fundus images. Starting with data collection, the algorithm follo ws the inte gration of fundal images and the disease labels to ensure preprocessing steps lik e resizing, normalization, and dataset partitioning. This f acili- ates the model training and v alidation. An emphasis is specically placed on the dataset imbalance to ensure an unbiase d model e v aluation. This research study has considered three CNN architectures namel y V GG16, InceptionV3, and ResNet50 for the feature e xtraction of the fundal images. GAP is implemented to simplify the comple x data from the images into a format that is easier to identify the patterns assosiated with the retinal diseases used in this study . T raining is conducted follo wed by e v aluation of the model’ s prediction ag ainst the established performance metrics. This w ay of structured comaparison and analysis of the results highlights the ef cienc y of the models in diagnosisng retinal diseases. The prediction also helps shed light into the details sur - ronding diabetic retinopath y , e xudation, and hemorrhagic retinopath y . W ith this rigorous process the research Pr edicting enhanced dia gnostic models: deep learning for multi-label ... (Sridhe vi Sundar ar ajan) Evaluation Warning : The document was created with Spire.PDF for Python.
58 ❒ ISSN: 2252-8938 helps to identify the promising CNN architecture for fundal images, therby of fering v aluable insights into the automated diagnosis of retinal conditions to help in adv ance opthalmic cure. Figure 2. Flo wchart for deep learning in ophthalmic disease classications 3. RESUL TS AND DISCUSSION Based on the comparison of the three architectures considered in the research we ha v e arri v ed at the follo wing analysis which re v eals the distinct performance features of V GG16, InceptionV3, and ResNet mod- els, highlighting their use in v arious applications based on their unique strength and weakness. The V GG16 model demonstrates superior performance in almost all e v aluated metrics. W ith a subset accurac y of 84.81%, it signicantly outperforms Inception and ResNet models in predicting the entire set of labels for an instance. This is complemented by its high precision, both on a macro (95.83%) and micro (95.56%) le v el, suggesting that the model is particularly ef fecti v e at minimizing f alse positi v es. Moreo v er , V GG16’ s recall rates (macro: 80.74%, micro: 79.63%) gi v es a strong indication about its capability to identify positi v e instances, which, combined with its precision, leads to the highest F1-scores among the models (macro: 87.22%, micro: 86.87%). These results highlight V GG16’ s balanced performance, making it a rob ust choice for tasks requiring precise label predictions. ResNet, on the other hand, presents itself as a competiti v e alternati v e, with a subset accurac y of 79.75%, positioning it between V GG16 and Inception in terms of the ability to match the full label set. While its precision and recall scores are moderately lo wer than those of V GG16, the y still sho wcase respectable performance (precision macro: 86.70%, rec all macro: 75.37%). The F1-scores for ResNet (macro: 80.00%, Int J Artif Intell, V ol. 14, No. 1, February 2025: 54–61 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 59 micro: 79.21%) further underscore its balanced precision-recall trade-of f, i ndicating its ef fecti v eness in a v ari- ety of scenarios where a slightly lo wer accurac y is acceptabl e. Inception’ s performance, while trailing behind V GG16 and ResNet, of fers insights into the challenges and limitations of the architecture within the specic task conte xt. W ith a subset accurac y of 66.67%, and the lo west precision (macro: 81.21%, micro: 78.05%) and recall (macro: 59.72%, micro: 60.38%) scores among the three, Inception appears to struggle more with correctly identifying positi v e instances and minimizing f alse positi v es. Its F1-scores (macro: 65.51%, micro: 68.09%) reect these challenges, suggesting areas for impro v ement, particularly in tasks requiring high preci- sion and recall. The T able 5 sho ws the results of the three CNN architecture models namely V GG16, Inception, and ResNet models carried out in this research study . T able 5. Comparati v e analysis of V GG16, Inception, and ResNet across k e y performance indicators Metrics V GG16 (%) Inception (%) ResNet (%) Subset accurac y 84.81 66.67 79.75 Precision (Macro) 95.83 81.21 86.70 Recall (Macro) 80.74 59.72 75.37 F1-Score (Macro) 87.22 65.51 80.00 Precision (Micro) 95.56 78.05 85.11 Recall (Micro) 79.63 60.38 74.07 F1-Score (Micro) 86.87 68.09 79.21 Hamming Loss 5.49 12.82 8.86 Jaccard Score (Macro) 78.62 53.10 69.30 Jaccard Score (Micro) 76.79 51.61 65.57 The graph representated in Figure 3 gi v es an o v ervie w between the performance metrics of the V GG16, Inception, and ResNet models across the k e y met ric as discussed in the T able 5. W e see that V GG16 emer ges as the most accurate a nd balanced model, of fering high precision and recall, which are critical for applications demanding stringent accurac y in label prediction. ResNet’ s performance, while not reaching the heights of V GG16, remains strong, suggesting its potential as a v ersatile model capable of deli v ering rob ust results across v arious tasks. Inception’ s lo wer performance metrics indicate potential dif culties in tasks re- quiring high precision and recall, highlighting the importance of model selection based on specic task re- quirements and the inherent trade-of fs between dif ferent architectures. This comparison not only emphasizes the strengths of V GG16 in achie ving high accurac y and a balanced precision-recall relationship b ut also illus- trates the v ersatility of ResNet and the tar geted applicability of Inception. The choi ce between these models should be guided by the specic requirements of the application, including the need for precision, recall, and the ability to correctly predict complete sets of labels. Figure 3. Grouped bar depicting the comparati v e analysis of V GG16, Inception, and ResNet across k e y performance indicators Pr edicting enhanced dia gnostic models: deep learning for multi-label ... (Sridhe vi Sundar ar ajan) Evaluation Warning : The document was created with Spire.PDF for Python.
60 ❒ ISSN: 2252-8938 4. CONCLUSION This study’ s ndings underscore the importance of model selection in the deplo yment of CNNs for image recognition tasks. The choice of model can signicantly inuence the outcome, depending on the specic requirements of accurac y , computational ef cienc y , and application conte xt. Future research should e xplore the adaptability of these models to emer ging image recognit ion challenges, including their performance on increasingly compl e x datasets and in real-w orld scenarios. Additionally , in v estig ating the inte gration of these architectures into ensemble methods could of fer paths to further enhance their accurac y and ef cienc y . In conclusion, while V GG16 emer ges as the top performer in our analysis, the strengths of ResNet and Inception in certain conte xts cannot be o v erlook ed. The decision to emplo y a particular model should be guided by the specic demands of the task at hand, balancing considerations of accurac y , ef cienc y , and practical constraints. As the eld of deep learning continues to e v olv e, so too will the capabilities and applicati ons of these models, promising ne w opportunities for inno v ation and impro v ement. REFERENCES [1] T . E. T an and T . Y . W ong, “Diabetic retinopath y: Looking forw ard to 2030, ” F r ontier s in Endocrinolo gy , v ol. 13, 2023, doi: 10.3389/fendo.2022.1077669. [2] P . G. P a v ani, B. Bisw al, and T . K. Gandhi, “Simultaneous multiclass retinal lesion se gmentation using fully automated RILBP-YNet in diabetic retinopath y , ” Biomedical Signal Pr ocessing and Contr ol , v ol. 86, 2023, doi: 10.1016/j.bspc.2023.105205. [3] P . Fu et al. , “Ef cac y and safety of pan reti nal photocoagulation combined with intra vitreal anti-VEGF agents for high- risk proliferati v e diabetic retinopath y: A systematic re vie w and meta-analysis, ” Medicine , v ol. 102, no. 39, 2023, doi: 10.1097/MD.0000000000034856. [4] M. Krichen, “Con v olutional neural netw orks: A surv e y , ” Computer s , v ol. 12, no. 8, 2023, doi: 10.3390/computers12080151. [5] A. Kay and M. Nguyen, “T ransfer learning with V GG16 deep con v olutional neural netw ork model ef fecti v ely dif ferentiates between subtypes of bright and dark lesions, ” In vestigative Ophthalmolo gy & V isual Science , v ol. 64, no. 8, pp. 242–242, 2023. [6] T . Castilla, M. S. Mart ´ ınez, M. Le gu ´ ıa, I. Larrabide, and J. I. Orlando, “ A ResNet is all you need: modeling a strong baseline for detecting referable diabetic retinopath y in fundus images, ” in 18th International Symposium on Medical Information Pr ocessing and Analysis , 2023, pp. 212–221, doi: 10.1117/12.2669816. [7] K. D. Bha v ani and M. F . Ukrit, “Design of inception with deep con v olutional neural netw ork based f all detection and classication model, ” Multimedia T ools and Applications , v ol. 83, no. 8, pp. 23799–23817, 2024, doi: 10.1007/s11042-023-16476-6. [8] S. P anchal et al. , “Retinal fundus multi-disease image dataset (RFMiD) 2.0: A dataset of frequently and rarely identied diseases, ” Data , v ol. 8, no. 2, 2023, doi: 10.3390/data8020029. [9] M. Im an, H. R. Arabnia, and K. Rasheed, “ A re vie w of deep transfer learning and recent adv ancements, ” T ec hnolo gies , v ol. 11, no. 2, 2023, doi: 10.3390/technologies11020040. [10] A. D. V airamani, “Detection and diagnosis of disease s by feature e xtraction and analysi s on fundus images using deep learning techniques, ” in Computational Methods and Deep Learning for Ophthalmolo gy , 2023, pp. 211–227, doi: 10.1016/B978-0-323- 95415-0.00009-7. [11] S. Guefrachi, A. Echtioui, and H. Hamam, “ Automated diabetic retinopath y screening using deep learning, ” Multimedia T ools and Applications , v ol. 83, no. 24, pp. 65249–65266, 2024, doi: 10.1007/s11042-024-18149-4. [12] S. Pra vin, S. Kanag asabapath y , V . Si v araman, S. Jayaraman, and P . Manicka v elu, “Ef cient CNN based detection of diabetic retinopath y , ” AIP Confer ence Pr oceedings , v ol. 2829, no. 1, 2023, doi: 10.1063/5.0156753. [13] K. A. He ger and S. M. W aldstein, “ Articial intelligence in retinal imaging: current status and future prospects, ” Expert Re vie w of Medical De vices , v ol. 21, no. 1–2, pp. 73–89, 2024, doi: 10.1080/17434440.2023.2294364. [14] O. Sri v asta v a, M. T ennant, P . Gre w al, U. Rubin, and M. Seamone, “ Articial intelligence and machine lea rning in ophthalmology: A re vie w , ” Indian J ournal of Ophthalmolo gy , v ol. 71, no. 1, pp. 11–17, 2023, doi: 10.4103/ijo.IJO 1569 22. [15] Z. Li, M. Xu, X. Y ang, Y . Han, and J. W ang, “ A multi-label detection deep learning model with attention-guided image enhancement for retinal images, ” Micr omac hines , v ol. 14, no. 3, 2023, doi: 10.3390/mi14030705. [16] P . Kaur and R. K. Singh, “ A re vie w on optimization techniques for medical i mage analysis, ” Concurr ency and Computation: Pr ac- tice and Experience , v ol. 35, no. 1, 2023, doi: 10.1002/cpe.7443. [17] R . Cha v an and D. Pete, “ Automatic multi-diseas e classication on retinal images using multile v el glo ww orm sw arm con v olutional neural netw ork, ” J ournal of Engineering and Applied Science , v ol. 71, no. 1, 2024, doi: 10.1186/s44147-023-00335-0. [18] H. Gao et al. , “Optimum design of a reusable spacecraft launch system using electromagnetic ener gy: an articial intelligence GSO algorithm, ” Ener gies , v ol. 16, no. 23, 2023, doi: 10.3390/en16237717. [19] A. R. Chłopo wiec, K. Karano wski, T . Skrzypczak, M. Grzesiuk, A. B. Chłopo wiec, and M. T abak o v , “Counteracting data bias and class imbalance—to w ards a useful and reliable retinal disease recognition system, ” Dia gnostics , v ol. 13, no. 11, 2023, doi: 10.3390/diagnostics13111904. [20] E. Hassan et al. , “Enhanced deep learning model for classication of retinal optical coherence tomograph y images, ” Sensor s , v ol. 23, no. 12, 2023, doi: 10.3390/s23125393. [21] M. S. Jamil, S. P . Banik, G. M. A. Rahaman, and S. Saha, “ Adv anced GradCAM++: Impro v ed visual e xplanations of CNN decisions in diabetic retinopath y , ” in Computer V ision and Ima g e Analysis for Industry 4.0 , 2023, pp. 64–75, doi: 10.1201/9781003256106-6. [22] T . Krzywicki, P . Brona, A. M. Zbrzezn y , and A. E. Grzybo wski, “ A global re vie w of publicly a v ailable datasets containing fundus images: Characteristics, barriers to access, usability , and generalizability , ” J ournal of Clinical Medicine , v ol. 12, no. 10, 2023, doi: 10.3390/jcm12103587. [23] N. Anton et al. , “Comprehensi v e re vie w on the use of articial intelligence in ophthalmology and future research directions, ” Dia gnostics , v ol. 13, no. 1, 2022, doi: 10.3390/diagnostics13010100. Int J Artif Intell, V ol. 14, No. 1, February 2025: 54–61 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 61 [24] A. K umar et al. , “ A no v el deep learning approach for retinopath y prediction using multimodal data fusion, ” International J ournal of Intellig ent Systems and Applications in Engineering , v ol. 12, no. 11, pp. 70–77, 2024. [25] D. Restrepo et al. , “Ophthalmology optical coherence tomograph y databases for articial intelligence algorithm: a re vie w , ” Seminar s in Ophthalmolo gy , v ol. 39, no. 3, pp. 193–200, 2024, doi: 10.1080/08820538.2024.2308248. [26] P . Ruamviboonsuk, N. Kaothanthong, V . Ruamviboonsuk, and T . Theeramunk ong, “T ransfer learning for articial intelligence in ophthalmology , ” in Digital Eye Car e and T eleophthalmolo gy , Cham: Springer International Publishing, 2023, pp. 181–198, doi: 10.1007/978-3-031-24052-2 14. BIOGRAPHIES OF A UTHORS Sridhe vi Sundararajan is acti v ely eng aged in her postgraduate studies, pursuing a Master of T echnology in Engineering Design at Symbiosis Institute of T echnology in Pune, India. She has around 15 years of e xperience in IT domain with C++, Unix being her main forte. Her area of interest is research in medical domain. She lik es to spend her time reading books related to medical e ld and e xploring unkno wn. She can be contacted at email: sride visundarajan@gmail.com. Harikrishnan Ramachandran is currently w orking as associate professor in the Depart- ment of Electronics and T elecommunication Engineering, Symbiosis Institute of T echnology , Pune Campus, Symbiosis International Deemed Uni v ersity , Pune, India. His main research interest in- cludes smart grid, optimization, internet of things, articial intelligence, and data analytics. He is an IEEE senior member , fello w life member of Institution of Engineers India, fello w life member of IETE India, life member of Indian Society for T echnical Education and life member of Computer Society of India. He can be contacted at email: dr .rhareish@gmail.com. Harshita Gupta is currently in the nal year of her bachelor’ s de gree in Electronics a nd T elecommunication Engineering from Symbiosi s Institute of T echnology , Pune, Sym biosis Interna- tional Deemed Uni v ersity . Additionally , she is pursuing a minor in Data Science from the Department of Computer Science of the same uni v ersity . W ith a k een interest in the dynamic elds of articial intelligence, machine learning, deep learning, and computer vision, she is poised to contrib ute to the future of technology . She can be contacted at email: harshitag2810@gmail.com. Y ashraj P atil is a nal year research scholar at Symbiosis Institute of T echnology , Pune Campus, Symbiosis International Deemed Uni v ersity , Pune, India, pursuing a Master of T echnology in Engineering Design. His main research interes t includes internet of things, articial intelligence, data analytics, and earth and space science. He is a UNIT AR Global Diplomac y Fello w , distinguished speak er at UN General Assembly Science Summit, GLOBE Scientist at the N ASA ’ s GLOBE Program and a peer mentor in N ASA ’ s SEES Summer Internship Program. He can be contacted at email: yashrajpatil2798@gmail.com. Pr edicting enhanced dia gnostic models: deep learning for multi-label ... (Sridhe vi Sundar ar ajan) Evaluation Warning : The document was created with Spire.PDF for Python.