ESTRO 2024 - Abstract Book

S637

Clinical - Breast

ESTRO 2024

Fibrosis after whole breast Radiotherapy is a relevant issue, impacting the quality of life of a not negligible number of patients. A dose/volume-response relationship for identified structures could drive the decision toward more conservative irradiation, including PBI, margin reductions, or hormonal treatment. Most studies consider PTV/CTV volume, but the manifestation of chronic symptoms is not only related to the subcutaneous tissue. The consequentiality between acute and late tissue alterations allows us to interpret the parenchymal alteration of the first layers of the skin, including the epidermal shell, papillary dermis and dermis with tufts and microvessels, and the hypodermis containing the first subcutaneous fat layers as potentially responsible for the late degeneration [1].

Our study considers the association between skin DVH and late symptoms related to fibrosis, investigating clinical and dosimetric factors to assess quantitative dose-response relationships for the breast skin.

Material/Methods:

The study population comprises 1325 consecutive early-stage patients who underwent conservative surgery followed by RT between 2009 and 2017. Patients received 40 Gy in 15 fractions without a dose boost. A median number of 4 segments was used within a tangential 3DCRT technique to obtain homogeneous dose distribution within PTV (hotspots <108%, V95%>95%). Follow-up visits were scheduled for 6-18-30-42-54-66 months, and the late effects were assessed through the SOMA/LENT scale. We focused on the dose distribution of the skin, defined as the first 5 mm body layers. According to previous studies, we selected such thickness to follow the tissue alteration measured in the epidermis, dermis, and hypodermis [1-3].

For the analysis, we considered as an event patients experiencing at least one of the following late symptoms: fibrosis, atrophy, telangiectasia, and pain (resulting in the comprehensive endpoint FATP).

We confirmed an increased incidence of FATP with time [1, 4]. We selected 42 months FU as a compromise between the number of events and the patients with complete FU at that time point. Moreover, due to the large population size, we preferred to analyse a moderate-severe grade for toxicity endpoints (i.e. FATP G2+). The association of DVH parameters (measured in cc) and clinical factors with the endpoint was tested through a machine-learning pipeline, giving a multifactorial logistic model as output. A preprocessing was used to reduce the dataset. Furthermore, we followed a novel methodology [5] to convert any predictors into the primary dosimetric variable identified by the model and apply personalised dosimetric constraints in the clinic.

Results:

In total, 1066/1325 patients could be considered for the analysis. We excluded patients with incomplete dosimetric information, bilateral breast treatment, and partial FU.

The rate for FATP G2+ was 3.8% (40/1066). After preprocessing, we selected 11/39 features. A 5-fold cross validation was applied to select the best variables. We used the Average Precision AUC (AUC-PR) as an optimisation metric, stressing the importance of correctly predicting the limited number of events. We iterated the procedure for increasing the number of variables to study the performance gain. The resulting optimal number of variables was 7; we decided to limit to 4 variables the complexity of the model due to the number of events resulting in V20 Gy, V42