Mukta, Jannatun NoorSiddiqui, Md. Saiful BariIslam, MD. RaisulEmon, Samir YeasirSumaiya, NowshinKhan, SakibLabib, Farhan2026-01-062026-01-0620252025-10ID 21301406ID 21301413ID 21301276ID 21301278ID 21301238http://hdl.handle.net/10361/27401Cataloged from PDF version of thesis.Includes bibliographical references (pages 99-108).This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science and Engineering, 2025.Diabetic Retinopathy, a silent threat to vision, is one of the major causes of vision impairment worldwide, where the retina of the eye is damaged before noticeable symptoms appear. Accompanying DR (Diabetic Retinopathy), DME (Diabetic Macular Edema) frequently develops, stating both are overlapping ocular conditions threatening visual acuity that can be effectively diagnosed by analyzing retinal images. However, relying only on a single modality has proven inadequate accuracy in distinguishing between DME and DR. Traditional diagnostic methods are employed primarily on fundus imaging, OCT (Optical Coherence Tomography), or OCTA (Optical Coherence Tomography Angiography). To date, single modality alone fails to provide a complete contextual understanding necessary for precise classification.This work proposes to offset the limitation by developing deep learning architectures that leverage several image modalities to improve classification performance and yield context-aware outputs. Specifically, the work proposes to develop personalized Convolutional Neural Networks (CNNs) driven mainly by superior fusion methods such as Multi-Head Self-Attention (MSA) Fusion, Gated Fusion, and Feature-wise Linear Modulation (FiLM) Fusion, with model interpretability at each step. The multimodal DR and DME classification strategy proposed architecture fuses two forms of image data or biomarkers so that the model may accommodate both structural and context-specific differences. Our proposed architecture has achieved an impressive accuracy of 95.52% and an F1-score of 0.975, outperforming the existing benchmark. Furthermore, this accuracy is achieved with a lower parameter count of 1.75 million and 2.57 million, with faster inference times of 19.289 ms and 19.843 ms for the two architectures, respectively, setting a state-of-the-art benchmark in the medical field.119 pagesenBRAC University theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission.Diabetic retinopathyDiabetic macular edemaFundus photographyFeature-wise linear modulationCNNsConvolutional neural networksSDG 3OCTAOptical coherence tomography angiographyDiabetic retinopathy--Diagnosis.Macula lutea--Diseases.Edema.Retina--Diseases--Imaging--Data processing.Diagnostic imaging.Data mining.Optical coherence tomography.Diabetic retinopathy--Classification.Retina--Tomography.Neural networks (Computer science).Thesis