Cervical Cancer Detection

Existing cervical cancer screening methods suffer from CNN compression penalties and malignancy-representation gaps, making them unreliable for clinical deployment, especially on diverse multi-center datasets.

Developed an Attention-Based Multiple Instance Learning (AB-MIL) pipeline on a pretrained ResNet50 backbone, processing uncompressed 1280×1280 patches with 128-dimensional gated attention and Focal Loss.

• 01Digitized and curated a multi-center clinical dataset from Bangladeshi hospitals
• 02Applied Cleanlab confident learning and predictive Shannon entropy for automated label-quality assessment
• 03Refined 3,601 initial images into 2,374 high-quality training bags via smart undersampling
• 04AB-MIL architecture on ResNet50 with 128-dim gated attention and Focal Loss
• 05Pixel-level diagnostic attention heatmaps from MIL spatial weights for clinical interpretability

• →95% accuracy on 2-class triage
• →92% accuracy on 3-class severity grading
• →Zero high-grade dysplastic cases misclassified as normal (fail-safe clinical behavior)
• →Trained on RTX 4080 Super

Clinical ML requires not just accuracy but interpretability and fail-safe behavior. The attention heatmaps proved more valuable than raw metrics for clinician trust.