ESTRO 2024 - Abstract Book

S4467

Physics - Machine learning models and clinical applications

ESTRO 2024

to disease-free survival (DFS) and overall survival (OS) of tumour infiltrating lymphocytes (TILs) for rectal cancer patients receiving radiotherapy using data from a phase III clinical trial.

Material/Methods:

(1) Data Source:

ARISTOTLE (ISRCTN09351447) [1] is a phase III clinical trial investigating whether the addition of irinotecan to the standard neoadjuvant chemoradiation (nCRT) of long-course capecitabine and radiotherapy improves outcomes for MRI-defined locally advanced rectal cancer patients. In this study, we used the clinical data and the digitised Haematoxylin-and-eosin-stained whole slide images (WSIs) from formalin-fixed paraffin-embedded samples of a sub cohort of patients whose DFS, OS and pre-treatment biopsies are available (N=414/589). The post-treatment resection samples are available for 195/414 patients.

(2) Quantifying the TIL density:

We designed an AI framework to automatically quantify the TIL density within the WSIs. First, the tissue type of each tile (224 pixels ×224 pixels with a pixel size of 0.5 µm) within the WSI was classified by a pre-trained EfficientNet [2], after which the tumour-associated regions were identified according to the tumorous tile distribution. Second, the TILs were detected by a pre-trained HoVerNet [3] within the targeted region. The EfficientNet was trained on the NCT CRC-HE-100K dataset [4] consisting of nine types of colorectal tissue. The HoVerNet was trained on the Lizard dataset [5] with five types of nuclei.

The TIL density was calculated using the equation:

TIL density = (N lym / N TUM ) ∗ (A 10HPFs / A tile ),

where N lym and N TUM were the number of lymphocytes and tumorous tiles. The A 10HPFs and A tile were the areas of 10 high power fields (2.4 mm 2 ) and a single tile (112 µm × 112 µm).