Application of Artificial Neural network in prediction of hypothyroidisim following radiation therapy and its comparison with conventional radiobiological models

Abstract

Purpose of this study, To determine the dose-response relationship of the thyroid for radiation-induced hypothyroidism in breast cancer patients whose supraclavicular is irradiated, according to 2 normal tissue complication probability models and artificial neural network, and to find the best-fit parameters of the models. Method and material: Clinical data and dose-volume histograms of 52 patients with breast cancer treated with 3DCRT were prospectively analyzed. Patients' serum samples (TSH, FT4, FT3) were measured before radiation therapy and also at a regular time interval until 1 year (every 3 months) after the completion of radiation therapy. Patients’ Thyroid blood tests before radiotherapy were considered as a basis for evaluation of the next blood test results. Thyroid DVHs were converted to 2 Gy/fraction equivalent doses using the linear-quadratic formula with ⍺/β = 3 Gy.Using two radiobiological models LKB, Log-logistic and also artificial neural network were used to calculate the NTCP of the thyroid gland in breast cancer patients. Performance and ranking of radiobiological models were performed using area under curve (AUC) and WAIC information index, respectively. Also, the function and correctness of the artificial neural network were determined using MATLAB and the values of MSE and R. The parameters of the models were obtained using the maximum probability estimation method by fitting the models with the clinical data of the patients.The models were fitted in a Bayesian setting and compared according to the Widely Applicable Information Criterion (WAIC). Result: Twenty-one (40%) out of 52 breast cancer patients developed RHT at a follow-up of 1 year after the end of radiation treatment. The mean dose to the thyroid gland in the patients was 18.24 Gy. The mean NTCP calculated with LKB and Log-logistic models was 49.87% and 49.23%, respectively. the highest-ranking for the Log-Logistic model was shown by the WAIC values, and then, the LKB model ranked as the next model. Also according to the AUC, the Log-logistic model has a higher performance. In the artificial neural network, the error rate was 0.146 for the training part and 0.128 for the test part. The mean squared squares were 0.3937 for training and 0.556 for test. The mean D50 parameter estimated from the models for the thyroid gland was 31/61 Gy.