Diagnosing Epidermal basal Squamous Cell Carcinoma in High-resolution, and Poorly Labeled Histopathological Imaging

The most appropriate method to uncover patterns from clinical records for each patient record is to create a bag with a variety of examples in the form of symptoms. The goal of medical diagnosis is to find useful ones first and then map them to one or more diseases. Patients are often represented as vectors in some aspect. Pathologists and dermatopathologists diagnose basal cell carcinomas (BCC), one of the most frequent cutaneous cancers in humans, on a regular basis. Improving histological diagnosis by producing diagnosis ideas, i.e. computer-assisted diagnoses, is a hotly debated research topic aimed at improving safety, quality, and efficiency. Due to their improved performance, machine learning approaches are rapidly being used. Typical images obtained by scanning histological sections, on the other hand, frequently have a resolution insufficient for today's state-of-the-art neural networks. Furthermore, weak labels hamper network training because just a small portion of the image signals the disease class, while the majority of the image is strikingly comparable to the non-disease class. The goal of this work is to see if attention-based deep learning models can detect basal cell carcinomas in histological sections and overcome the ultra-high resolution and poor labeling of full slide images. With an AUC of 0.99, we show that attention-based models can achieve nearly flawless classification performance.