A Semantic Quality-Gated AI Framework for Reliable Histopathological Diagnosis

Digital histopathology is gaining importance in computer-aided disease diagnostics. The accuracy of artificial intelligence systems may be lowered due to a variety of problems related to tissue processing, staining, blurriness, folds, dirt, and other imaging artifacts. Poor images may spoil the results of diagnostic modeling. Therefore, this study is focused on the development of a quality-checking system for histopathological research that defines the quality of the slides before the disease is analyzed. To achieve this, a digital image of the prepared tissue slide is taken and analyzed for quality. The approach serves to locate the area of interest using U-Net and to identify artifacts based on ResNet-50 processing. Next, the regions of tissue and artefact are compared to check if the identified artefact has changed the tissue relevant for diagnosis. The outcome of this comparison is the decision whether the slide is accepted or not. The accepted slides then undergo high power analysis or imaging and their staining is normalized by either Macenko or Reinhard approach. The normalized images are processed for preliminary disease detection using ResNet-50 and Vision Transformer. U-Net++ then isolates the area of disease followed by morphological characteristics extraction like tumor size, mitotic index and gland irregularity. The combined features are then processed along with deep features available for severity determination and disease classification using machine learning techniques like Random Forest and XGBoost. The tool is also able to generate report in form of templates and store the analyzed case in cloud storage.