ESTRO 2024 - Abstract Book

S3164

Physics - Autosegmentation

ESTRO 2024

not limited to: lesion location within lung lobes, shape (round vs oval vs complex), peri-tumor emphysema and the presence of peripheral glass opacities (pGGOs). Two classification models were built, considering the Dice coefficient as either a binary (< or ≥0.8) or continuous outcome (linear Model1 and factor regression model 2, respectively). A train:test ratio of 0.7 and a cross-validation of 10 were used in both cases. Analyses were performed with R libraries caret and glmnet.

Results:

Considering the 70 available CTs, manually-segmented lesions had a higher volume than their U-net counterparts: 4.6 (IQR 2.0-12.2) vs 3.2 (IQR 1.3-10.3) cm3, with a median percentage difference of 21.9%. Median dice coefficient was 0.80 (IQR 0.71-0.86). Model1 showed that lesion shape and pleuric contact were the main determinants of performance, with complex shape being associated with Dice values <0.8 (Fig. 1a). Conversely Model2 suggest the importance of additional variables such as the proximity to lung fissures, intra-lesion density (solid vs subsolid), lesion margins (regular vs irregular) and the presence of pGGOs (Fig 1b).

1a