Transformer-based framework for multi-class segmentation of skin cancer from histopathology images

Muhammad Imran; Mohsin Islam Tiwana; Mashood Mohammad Mohsan; Norah Saleh Alghamdi; Muhammad Usman Akram

doi:10.3389/fmed.2024.1380405

Transformer-based framework for multi-class segmentation of skin cancer from histopathology images

Front Med (Lausanne). 2024 Apr 29:11:1380405. doi: 10.3389/fmed.2024.1380405. eCollection 2024.

Authors

Muhammad Imran¹, Mohsin Islam Tiwana¹, Mashood Mohammad Mohsan², Norah Saleh Alghamdi³, Muhammad Usman Akram²

Affiliations

¹ Department of Mechatronics Engineering, National University of Sciences and Technology, Islamabad, Pakistan.
² Department of Computer and Software Engineering, National University of Sciences and Technology, Islamabad, Pakistan.
³ Department of Computer Sciences, Princess Nourah Bint Abdulrahman University, Riyadh, Saudi Arabia.

Abstract

Introduction: Non-melanoma skin cancer comprising Basal cell carcinoma (BCC), Squamous cell carcinoma (SCC), and Intraepidermal carcinoma (IEC) has the highest incidence rate among skin cancers. Intelligent decision support systems may address the issue of the limited number of subject experts and help in mitigating the parity of health services between urban centers and remote areas.

Method: In this research, we propose a transformer-based model for the segmentation of histopathology images not only into inflammation and cancers such as BCC, SCC, and IEC but also to identify skin tissues and boundaries that are important in decision-making. Accurate segmentation of these tissue types will eventually lead to accurate detection and classification of non-melanoma skin cancer. The segmentation according to tissue types and their visual representation before classification enhances the trust of pathologists and doctors being relatable to how most pathologists approach this problem. The visualization of the confidence of the model in its prediction through uncertainty maps is also what distinguishes this study from most deep learning methods.

Results: The evaluation of proposed system is carried out using publicly available dataset. The application of our proposed segmentation system demonstrated good performance with an F1 score of 0.908, mean intersection over union (mIoU) of 0.653, and average accuracy of 83.1%, advocating that the system can be used as a decision support system successfully and has the potential of subsequently maturing into a fully automated system.

Discussion: This study is an attempt to automate the segmentation of the most occurring non-melanoma skin cancer using a transformer-based deep learning technique applied to histopathology skin images. Highly accurate segmentation and visual representation of histopathology images according to tissue types by the proposed system implies that the system can be used for skin-related routine pathology tasks including cancer and other anomaly detection, their classification, and measurement of surgical margins in the case of cancer cases.

Keywords: deep learning; histology image analysis; image transformers; segmentation; semantic segmentation.

Grants and funding

The author(s) declare financial support was received for the research, authorship, and/or publication of this article. This research study was funded by the Princess Nourah Bint Abdulrahman University Researchers of supporting project number (PNURSP2024R40) and Princess Nourah Bint Abdulrahman University, Riyadh, Saudi Arabia.