ESTRO 2024 - Abstract Book

S3021

Physics - Autosegmentation

ESTRO 2024

The clinical significance of false positive (FP) and false negative (FN) prediction errors, depends on proximity to true positive (TP) regions. When FP pixels are close to a sufficiently large TP region, they will only highlight the same region to the operator and can be disregarded. Equally, if FN pixels are close to a large TP region, they can be disregarded. Geometric Distance Correction (GDC) accommodates this by reclassifying such pixels. The definition of ‘closeness’ depends on the size of the true positive region, to avoid bias from few-pixel TP regions.

Visual assessment allowed further validation of clinical value, especially where residual errors in gold-standard contours occurred, or uncertainty was correlated with low image contrast or artefact.

Results:

ACo showed excellent performance on uncertainty prediction for both internal and external segmentations, across all OARs except lenses. MCC was higher on IAS and lower-quality external segmentations (MRIu) than high-quality ones (MRIeMRI).

Visual assessment indicated excellent correlation between differences to gold-standard contours and predicted confidence maps [fig1].

On IAS and MRIu, average MCC (excluding lenses) varied from 0.6 to 0.9, while average FPR and FNR were ≤0.13 and ≤0.21, respectively. For MRIeMRI, average MCC (excluding lenses) varied from 0.4 to 0.8, while average FPR and FNR were ≤0.37 and ≤0.22, respectively. [fig2]