fdamptey/PUBLIC-HEALTH-INFORMATICS_PROJECT-WORK

This AI-powered breast cancer screening tool using ultrasound images and deep learning. Convolutional Neural Network achieves high accuracy and AUC, offering a scalable alternative to mammography for early detection in low-resource settings. # Exploring Deep Learning AI Ultrasound as a Primary Breast Cancer Screening Tool ## 👨‍🔬 By: **Frederick Damptey** **Benjamin Odoom Asomaning** ## Overview This project investigates the use of deep learning, specifically **transfer learning with convolutional neural networks (CNNs)**, to classify breast ultrasound images into **normal, benign, and malignant** categories. Using a carefully preprocessed public dataset and models like **EfficientNetB0**, we aim to provide a **low-cost, highly accurate screening tool** that can be deployed especially in **resource-limited settings**. ## Objectives - Fine-tune deep learning models for breast cancer classification. - Evaluate model performance using clinically relevant metrics: **accuracy, AUC, sensitivity, and specificity**. - Compare AI model performance against mammography AI and human readers from study findings. - Evaluate feasibility of using **deep learning AI-powered ultrasound** as a **primary screening tool**, especially in LMICs (low- and middle-income countries). ## Dataset - **Source:** Kaggle Breast Ultrasound Images Dataset (BUSI) The dataset consists of **780 B-mode ultrasound images** divided into three classes: - `normal` - `benign` - `malignant` Ground truths were established through comparison with mammograms and confirmed by histopathology. ## Methods ### Preprocessing - Resizing images to `224x224` - RGB conversion for model compatibility - Image normalization and augmentation - K-Fold Cross-Validation (K=5) ### Models Used All models were pretrained on ImageNet: - EfficientNetB0 **(Best Performing Model)** - ResNet50 - VGG16 - InceptionV3 ### Evaluation Metrics - **Accuracy** - **AUC (Area Under Curve)** - **Sensitivity** - **Specificity** - Confusion matrix and ROC curves I maintained all hyperparameters for consistency across all models. --- ## 🏆 Results | Model | Accuracy | AUC | Sensitivity | Specificity | |---------------|----------|-------|-------------|------ …