A deep learning based approach to semantic segmentation of lung tumour areas in gross pathology images

Gross pathology photography of surgically resected specimens is an often overlooked modality for the study of medical images that can provide and document useful information about a tumour before it is distorted by slicing. A method for the automatic segmentation of tumour areas in this modality could provide a useful tool for both pathologists and researchers. We propose the first deep learning based methodology for the automatic segmentation of tumour areas in gross pathological images of lung cancer specimens. The semantic segmentation models applied are Deeplabv3+ with both a MobileNet and Resnet50 backbone as well as UNet, all models were trained and tested with both a DICE and cross entropy loss function. Also included is a pre and post-processing pipeline for the input images and output segmentations respectively. The final model is formed of an ensemble of all the trained networks which produced a tumour pixel-wise accuracy of 69.7% (96.8% global accuracy) and tumour area IoU score of 0.616. This work on this novel application highlights the challenges with implementing a semantic segmentation model in this domain that have not been previously documented.