OsteoNeuralNetwork

Gallery

About

I'm a biomedical-to-CS student, and I wanted to see how far a genuinely explainable (Use of heatmapping) medical imaging prototype could get on near-zero infrastructure. It's a DenseNet-121 trained on BTXRD (3,746 radiographs), three classes: normal, benign, malignant. Malignant is only 9.1% of the data, so a model predicting "not malignant" scores 90.9% accuracy while missing every cancer. Focal loss with inverse-frequency alpha, thresholds picked on validation only, probabili...