ESTRO 2024 - Abstract Book

S3048

Physics - Autosegmentation

ESTRO 2024

Addressing this requires a combination of standardized delineation guidelines, continuous clinician training, and artificial intelligence-driven contouring software.

Focusing on the last aspect: accurate data and ground truth labeling are paramount to effective learning, acting as a training reference to uncover the correct anatomical structures and tumor boundaries. Yet, current research on target volume autosegmentation often falls short in performance due to label quality. We propose harnessing the potential of histopathology to provide ground-truth tumor extent to planning acquisitions. Indeed, these digitized post-surgery specimens represent the gold standard for tumor characterization by offering a comprehensive microscopic representation of disease invasion. Through an innovative automatic multimodal registration method, we can map and transfer accurate tumor contours from histology to planning CT images and leverage them to build a histology driven target volume autosegmentation model. This study involves a unique dataset of 180 HNC patients, including both a pre-operative CT scan and 5-12 digitized histopathology slides from laryngeal tissue surgery. Two junior and two senior radiation oncologists delineated the GTV on CT for a posteriori comparison, while two senior pathologists contoured the gold standard macroscopic tumor on histology. Our work relies on two pillars (Figure 1): - Capitalizing on a validated method that learns a mapping between digital pathology and CT scans [1]. Advanced deep-learning registration models are employed to address tissue collapse and shrinkage due to the histological process, ensuring accurate alignment and out-of-plane deformation handling between both modalities. The model serves as a bridge, enabling the transfer of tumor delineations from histology to CT. These histological contours are proven accurate and consistent across expert pathologists (dice = 96%) [2]. - Building a model leveraging these mapped histological contours to develop a precise target volume delineation model from CT. To this end, we trained a latent diffusion model, which is a leading method in deep learning with state-of-the-art performance in many tasks [3]. Starting from random noise and conditioning on the input CT, the model learns to progressively denoise the input and generate the desired segmentation mask. A unique aspect of diffusion models is their non-deterministic process: a slightly different sampling in the input noise can lead to different yet valid segmentation outputs and to the quantification of the inherent uncertainties in the delineation process [4]. In the context of RT, understanding these uncertainties aids clinicians in assessing segmentation reliability and improves subsequent treatment planning. Material/Methods: