KONGU ENGINEERING COLLEGE DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING MULTIMODAL AI ANEMIA DETECTION APPLICATON USING NAIL BED AND EYE FUSION ANALYSIS

[Audio] KONGU ENGINEERING COLLEGE DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING MULTIMODAL AI ANEMIA DETECTION APPLICATON USING NAIL BED AND EYE FUSION ANALYSIS PROJECT BY: PROJECT GUIDE: 22ECR207 – SWATHI M Dr. T. MEERA DEVI 22ECR211 – THAARINI S PROFESSOR /ECE 22ECR233 – VINUDHARAN M.

[Audio] INTRODUCTION Anemia is a blood disorder caused by low hemoglobin, leading to fatigue and weakness. Early detection is essential to prevent severe health complications and improve quality of life. Non-invasive detection using images of the eye conjunctiva and nail bed provides a faster, safer alternative to blood tests. Deep learning models can automatically extract features and classify images as anemic or healthy. Fusing results from conjunctiva and nail images may improve prediction accuracy in future implementations..

[Audio] PROBLEM STATEMENT Traditional anemia detection methods are invasive, time-consuming, and not accessible for routine screening. Single-source image detection (eye or nail) may have limited accuracy due to variations in lighting, skin tone, and image quality. There is a need to compare or integrate multiple image sources to improve detection reliability. A fast, automated, and non-invasive prediction system is required for real-time health monitoring..

[Audio] EXISTING APPROACH Laboratory Blood Tests (Traditional): Hemoglobin (Hb) measured from blood samples; highly accurate but invasive, time-consuming, and resource-dependent. Clinical Visual Examination: Doctors inspect palpebral conjunctiva, nailbed, and lips for pallor; non-invasive but subjective and dependent on lighting and clinician expertise.

[Audio] LITERATURE REVIEW S.NO TITLE AUTHORS NAME PUBLISHER JOURNAL NAME & YEAR METHODOLOGY INFERENCE 1. Iron deficiency anemia detection using machine learning models: A comparative study of fingernails, palm and conjunctiva of the eye images Justice Williams Asare, Peter Appiahene, Emmanuel Timmy Donkoh, Giovanni Dimauro Published in Engineering Reports in 2023, with DOI reference to eng2.12667 Collected conjunctiva, palm, and fingernail images. Preprocessed with ROI segmentation and CIELAB feature extraction. Trained models: CNN, Naïve Bayes, SVM, k-NN, Decision Tree; used data augmentation and cross-validation Non-invasive image-based ML detection is effective; CNN overall strongest, Naïve Bayes slightly better on palm data 2. Prediction of anemia in real-time using a smartphone camera processing conjunctival images Leon Zhao1, Alisa VidwansID2, Courtney J. Bearnot1, James Rayner1, Timmy Lin1, Janette Baird1, Selim Suner1 , Gregory D. JayID1 Public Library of Science (PLOS) PLOS ONE 2024 YOLOv5 was used to detect anemia from eye (conjunctiva) images using a custom-labeled dataset. The model showed high accuracy, proving deep learning's effectiveness for non-invasive anemia detection..

[Audio] S.NO TITLE AUTHORS NAME PUBLISHER JOURNAL NAME & YEAR METHODOLOGY INFERENCE 3. A Deep Learning-based System for Detecting Anemia from Eye Conjunctiva Images taken from a Smartphone Pallavi; Bijit Basumatary; Rahul Shukla; Rakesh Kumar; Bodhisatwa Das; Ashish Kumar Sahani ETE Technical Review, Vol. 41, Issue 3 (2024), published online 10 Aug 2023 U-Net segmentation (IoU = 0.922); Classification model (deep learning) trained on 300 images Effective smartphone-based, non-invasive anemia screening; high potential for low-resource settings via chatbot deployment 4. Convolutional Neural Network for Anemia Detection Based on Conjunctiva Palpebral Images Rita Magdalena; Sofia Saidah; Ibnu Da’wan Salim Ubaidah; Yunendah Nur Fuadah; Nabila Herman; Nur Ibrahim Jurnal Teknik Informatika (Jutif), 2022 (DOI: 10.20884/1.jutif.2022.3.2.197) CNN with 5 layers (filter sizes: 3×3, 5×5, 7×7, 9×9, 11×11; channels: 16→128), FC layer + sigmoid; dataset: 2000 images split 1440/160/400 Effective non-invasive anemia detection via CNN; high accuracy; suitable for real-time Android app deployment 5. Machine vision model using nail images for non-invasive detection of iron deficiency anemia in university students Navarro-Cabrera et al Frontiers in Big Data, Vol. 8, April 2025 909 nail images; smartphone camera; Rad-67 hemoglobin readings; preprocessing; models: DenseNet169, InceptionV3, Xception Promising non-invasive anemia screening in young adults; performance suggests scope for enhancement.

[Audio] S.NO TITLE AUTHORS NAME PUBLISHER JOURNAL NAME & YEAR METHODOLOGY INFERENCE 6. Artificial Intelligence for Anemia Screening, Diagnosis, and Clinical Management Damilola B. Olawade; Emmanuel O. Olatunji; Opeyemi A. Alabi; et al Biomedical Artificial Intelligence, Elsevier, 2025 Comprehensive review of AI-based non-invasive anemia detection systems using ensemble learning and multimodal fusion strategies. Discusses score-level (late) fusion, feature fusion, and decision fusion. Highlights weighted score fusion to integrate prediction probabilities from independent deep learning models to improve robustness and reduce bias. Concludes that score-level fusion significantly improves diagnostic reliability and generalization over single-modality models, especially under varying lighting and skin-tone conditions. Supports the use of eye–nail score fusion mechanisms for mobile and low-resource anemia screening applications 7. Lightweight Hybrid Model for Bone Fracture Detection Using MobileNetV2 Feature Extraction and Ensemble Learning Bahriye Isgor; Murat Koklu Journal of Future Artificial Intelligence and Technologies, Vol. 2, No. 3, Dec 2025 MobileNetV2 was used as a frozen deep feature extractor and combined with ensemble classifiers. Experiments were conducted on 9,463 augmented X-ray images using 10-fold cross-validation and standard performance metrics. Score-level fusion improves reliability and generalization over single-modality models, supporting eye–nail fusion for mobile anemia screening..

[Audio] PROPOSED METHOD Dual Capture: Acquires images of the conjunctiva and nail bed simultaneously. Parallel AI Processing: Two separate CNNs analyze the preprocessed images to get individual probability scores. Weighted Fusion: The individual scores are combined via a Weighted Averaging Calculation for a final, reliable score. Compares the final score to an ROC based Optimal Threshold to classify the result as "Anemic" or "Non-Anemic"..

[Audio] A New Artificial Intelligence Approach Using Extreme Learning Machine as the Potentially Effective Model to Predict and Analyze the Diagnosis of Anemia A New Artificial Intelligence Approach Using Extreme Learning Machine as the Potentially Effective Model to Predict and Analyze the Diagnosis of Anemia DATASET Classes Female Male Before Augmented After Augumented Conjunctiva eye 306 404 710 70400 Nailbed 306 404 710 70400.

[Audio] METHODOLOGY. 10. Start: User Initiate s Irnage Acquisition Capture Conjuct:iva Image using C am era Preprocessing Pipeline: Conjunctiva Center Region of Interest CroppingC730/0): Palpebral Tissue Resize Irnage to 224 pixels x 224 Pixel I scale) Data Augurnentati on t ati on D eep Lear-ning Mo delCEye): p r etrain e d) SE Attention Block F O Cal LOSS Model Output: P_Eye - Conjunctiva Anemic Probability S c ore Capture e d Ixnage using Preprocessing Pipeline: Nailbed age Center Region oflnterest Cropping(730/0): Nail Region Resize Irrxage to 224 pixels 224 Pixel Normalization(O— I scale) D ata Augurn entatio n ss Flip Deep Leaming il en etv2 arn ageN et pretr aine d) SE Attention Block F oc l_ngs Model Output: P_Nai1 - Nailbed Aner-nic Probability Score Score Fusion Weighted Averaging C alculation: Final Score = Weight_Eye x P_Eye + Weight_Nai1 x P_bTai1 Decision Logic: ROO -based optirnal Threshold selection 11 _ O — 11 9 Final Score Threshold Diagnosis Result: Anemic Harnoglobin S evet-it-y Range Cla s gific ation o O — 10.9 Final Score < Threshold An e rnic Display to User End Systetn Execution.

[Audio] MOBILENET V2 Mobile & Edge Deployment Designed for smartphones, IoT devices, and embedded systems due to low memory usage and fast inference. Real-Time Image Classification Widely used for classifying objects in images (e.g., people, medical features like nailbed or conjunctiva, everyday objects). Computer Vision Tasks Serves as a backbone for object detection, face recognition, and semantic segmentation models. Medical & Healthcare Applications Used in lightweight medical imaging apps for screening tasks where computational resources are limited. Transfer Learning Friendly Pretrained on ImageNet and easily fine-tuned for custom datasets, making it popular in research and industry..

[Audio] GRAD-CAM BASED MODEL VALIDATION Grad-CAM was applied to the final convolutional layers to visualize regions influencing anemia prediction. Eye model attention was concentrated on the palpebral conjunctiva, while nail model focused on the central nailbed region. ROI-based preprocessing reduced background activation and improved localization accuracy. Heatmaps confirmed that the model learned clinically relevant features rather than noise or artifacts. Grad-CAM validation enhances model interpretability, reliability, and clinical trust..

[Audio] WORKFLOW IN ANDROID APP Medical Image Acquisition Clinical images are captured using smartphones, dermatoscopes, slit-lamp cameras, or digital microscopes (e.g., nailbed, conjunctiva, skin lesions). Image Preprocessing Images are resized (usually 224×224), normalized, noise-reduced, and sometimes cropped or enhanced to highlight medical regions of interest. Feature Extraction using MobileNetV2 The preprocessed image is passed through MobileNetV2, which uses depth wise separable convolutions and inverted residual blocks to efficiently extract medically relevant features. Fine-Tuned Classification Layer Transfer learning is applied by retraining final layers on medical datasets to classify conditions (e.g., normal vs abnormal, disease categories). Prediction & Clinical Output The model outputs probabilities, risk scores, or diagnostic suggestions, which are displayed in the app for screening, monitoring, or decision support. Android Studio Download Free - 2025.2.2.7 | TechSpot.

[Audio] THANK YOU. 14. THANK YOU.