The MuTILs model design involves joint classification of tissue regions and cell nuclei using a panoptic segmentation algorithm. It consists of two parallel U-Net models for segmenting tissue regions and nuclei at different magnifications. The model shares information between tissue region and nucleus segmentation to provide context-aware assessment of tumor infiltrating lymphocytes. The PanopTILs dataset was created by combining annotations from two public datasets and is used for panoptic segmentation of tissue regions and cell nuclei. The MuTILs model was trained using this dataset and additional annotations from the Cancer Prevention Study II cohort. The model’s performance was validated using internal-external cross-validation and compared to pathologist visual TIL scoring.
For computational TIL score calculation, a tiling procedure is used to analyze informative tiles at high resolution. Different TIL score variants are calculated using aggregation strategies and saliency-weighted averaging of informative tiles. A linear calibration is applied to scale computational scores to match visual scores. Clinical outcomes analysis is performed using progression-free interval as the endpoint, with Kaplan-Meier curves examined for patient subgroups based on TIL score thresholds. The study adhered to ethical regulations including obtaining informed consent from participants.
In summary, the MuTILs model design and training process, along with the creation and utilization of the PanopTILs dataset, were described. The computational TIL score calculation and calibration, as well as the clinical outcomes analysis, were also outlined in the content.
Source link
Source link: https://www.nature.com/articles/s41523-024-00663-1