DIGITAL HISTOPATHOLOGY AI

The Digital Histopathology project funded by an HEC-NRPU grant helped us test and benchmark our low-cost, low-resolution (LCLR) approach to digital histopathology that uses 5MP field view images instead of whole-slide images (WSI). Our team of clinicians, and ML/DL engineers, along with the team of histopathologists at the Rehman Medical Institute (RMI) developed a pipeline for the procurement of biopsy slides, digitisation of metadata, slide labelling, image acquisition and image storage. The first images for four cancers (breast, colon, oral, and gastrointestinal) was launched as HistoVault v1, with close to 18000 images (including both WSI and LCLR). The team trained and tested seven different AI models ultimately selecting and fine-tuning a hybrid TransUnet model for binary and multiclass classification of all four cancers.

Histovault: Pakistan’s Largest Histopathology Imageset

Comprehensive dataset of 17,980+ histopathology images across multiple resolutions enabling low-cost cancer detection.

Curated 480 slides (240 cancer, 240 normal) across oral, gastric, colorectal, and breast cancers, with perfect balance in each cancer type (60 cancer/60 normal).

Digitized at three different resolutions: gold-standard Whole Slide Imaging (10,000 MP), low-cost 10X (5 MP), and ultra-low-cost 40X magnification (5 MP).

Cost comparison: standard WSI acquisition at PKR 350 versus LCLR approach at just PKR 2.92 per image; storage requirements reduced from 900MB to 4MB per image.

Comprehensive annotation including cancer type, specific diagnosis, anatomical site, and grade, making this Pakistan's largest labeled histopathology dataset with 17,980 images.

TransUNet: High-Accuracy Cancer Detection on Low-Cost Images

Aim: A hybrid deep learning architecture integrating U-Net and Transformer modules to achieve clinical-grade accuracy in segmenting low-resolution histopathology images

TransUNet (96.8%) and EfficientNet (97.26%) demonstrate exceptional classification performance with minimal loss rates (3% and 20% respectively) on low-cost images.

Dramatic efficiency gains: processing time reduced from 8 minutes to 3 seconds per image; storage requirements cut from 900MB to 4MB per image.

User-friendly interface enables immediate diagnosis: pathologists upload images through simple file selector and receive instant cancer probability assessment with 96.8% accuracy.

Mobile deployment underway: ultra-low-resolution (4x) smartphone-captured histology analysis could revolutionize diagnosis in remote areas, bringing AI-powered cancer detection to resource-limited settings across Pakistan.