Halo ai

To classify tumor tissue into ‘stroma’, ‘epithelium’, and ‘artifacts’ compartments (Figure 2), we trained the HALO^® AI (Indica Labs, Albuquerque, NM, USA) Densenet v2 classifier using manual annotations provided by the pathologist (JD) in BCG-treated patient cohort. The annotations were created using a built-in HALO^® AI annotation tool via the user-friendly graphical user interface. The image data within the annotated regions are automatically incorporated into the model training pipeline through HALO^® AI’s native methods, eliminating the need for manual data management. The ‘artifacts’ class was incorporated to exclude areas of coagulation, necrosis, hemorrhage, or calcifications that could potentially interfere with further analyses. The quality of tissue classification masks produced by HALO^® AI was visually assessed by a pathologist (JD). Following initial tissue classification and artifact exclusion, due to the very low area (less than two mm²) of the remaining tumor, 9 cases were excluded, leaving 157 for further analyses. The clinical and pathological data of these patients are summarized in Table 1.

A pathologist (RS) reviewed all archived slides to identify the most representative formalin-fixed paraffin-embedded (FFPE) block. The selected block includes non-necrotic HCC tissue, and, whenever possible (in 103 samples, accounting for 98.1% of the cases), the surrounding peritumoral liver parenchyma. The 3 µm sections were stained using a modified Gordon and Sweet’s silver impregnation protocol combined with Picric Acid–Sirius Red, referred to as GSPS (see Figure 2E–H, supplementary Table S1). This method is standard for liver and bone marrow samples at the National Center of Pathology. Throughout this paper, we define the red-staining fibers in the thick fibrous septae as ‘collagen’, and the delicate black linear strands mostly located in the epithelial areas as ‘reticulin’.
All slides were subsequently digitized at 20× magnification (0.5 μm per pixel) using an Aperio^® AT2 DX scanner (Leica Aperio Technologies, Vista, CA, USA). A pathologist (RS) reviewed the images to mark the malignant (HCC) and non-malignant (peritumoral liver) areas on each slide by placing manual annotations (see Figure 3A,D). A HALO^® AI (Indica Labs, Albuquerque, NM, USA) classifier was used to categorize the tissue into hepatocytes (indiscriminately malignant and non-malignant), stroma, and background classes.