Laryngeal Cancer Lesion Segmentation in P63 Immunohistochemically Stained Histology Images

Abstract

In 2017, 1.09 million active laryngeal cancer cases were reported globally, increasing 23.8% in the last three decades. Hematoxylin and eosin (H&E) stained histological analysis is used as a gold standard for diagnosing and staging H&N cancers. Recent studies showed the effectiveness of P63 stain along with H&E stain. Histological analysis of H&E and P63 stained images is performed manually by a pathologist in current medical practices. This process is tiresome and requires specialized skills. The primary factor that the process is manual is that P63 staining is a new technique, which has been developed recently. The second factor is the lack of an available dataset in the recent past. We present an automatic system for laryngeal cancer lesion segmentation in H&E and p63 stained images. The proposed method is an encoder-decoder-based UNet++. We have used MEDISP HICL dataset for training and testing of the proposed network. To our knowledge, only one attempt for this purpose exists in the literature. The proposed network has shown superior accuracy, dice score, and MeanIoU, 95.37%, 87.82, and 0.7868, respectively.