ESTRO 2024 - Abstract Book

S4511

Physics - Machine learning models and clinical applications

ESTRO 2024

Finally, ilium, lumbar spine, pelvic bone, bone (ilium + lumbar spine + pelvic bone) and subregion mean doses were computed for every patient. They were then used to predict haematological toxicity using a logistic regression model (5-fold cross-validation, 1000 repetitions).

Results:

Mean Dice scores of 0.95±0.02, 0.94±0.04, 0.96±0.02 and 0.96±0.01 were obtained for the non-rigid registration towards the common coordinate system of the ilium, the lumbar spine, the pelvic bone and the bone, respectively. Mean values of 0.95±0.02, 0.95±0.02, 0.95±0.05 and 0.95±0.02 were obtained for the non-rigid registration back to every patient native space of the same structures.

The anatomical subregions found to be associated with the two haematological toxicity endpoints are shown in Figure 1.

Figure 1. Anatomical subregions predictive of haematological toxicities. Subregions were identified after a permutation test on every voxel of the planned dose distribution between patients with and without toxicity. They correspond to clusters of voxels with a p-value < 0.01. Two endpoints, based on (a) the absolute lymphocyte count and (b) the absolute neutrophil count, were considered.

Median mean doses of patients with and without toxicity for the two endpoints are presented in Table 1. Subregion mean doses of patients with toxicity were significantly higher, according to a Mann-Whitney U test.

ALC-based endpoint

ANC-based endpoint

No toxicity

Toxicity

No toxicity

Toxicity

Ilium

28.76

29.61 ns

28.62

29.20 ns

Lumbar spine

34.49

36.26 ns

35.03

34.65 ns

Pelvic bone

31.42

32.61 ns

31.42

31.84 ns

Bone

31.71

32.45 ns

31.78

32.22 ns

Subregion

45.69

48.15 **

31.61

35.16 **

Table 1. Median mean doses (in Gy) of patients with and without toxicity.