Smart Microscope for Online Detection of Hepatocellular Carcinoma

A common goal of microscopy is to assist experts in analyzing the sample status and tailoring research strategies by observing and collecting specific images. Utilizing microscopy to examine the pathological lesions of tissue slices for Hepatocellular Carcinoma (HCC) and provide treatment plans is a common practice among pathologists. Recently, pre-trained artificial intelligence (AI) models have been applied to assist pathological whole slide images (WSIs) detection. However, most analysis methods rely on digital WSIs from a scanner. This dependence limits the sensitivity of models due to image style differences and increases imaging and computational costs. To address these defects, we propose a smart microscope for executing online HCC detection. By integrating a fine-tuned vision-language model (VLM), the framework achieves real-time imaging and detection for morphological correlations of liver tissues under the lens. We fine-tuned the VLM by 365 collected WSIs containing HCC and evaluated the proposed method on 87 clinically hematoxylin and eosin (H&E) stained histological slides. The experiments demonstrated that our microscope achieved 100% sensitivity during online detection, which consistency is in pair with pathologists. Meanwhile, the high-confidence HCC fields are recorded precisely. The proposed framework aims for pathologists to improve diagnostic efficiency.