Session Item

Physics track: Radiation protection, secondary tumour induction and low dose
Poster
Physics
09:17 - 09:25
Transfer learning and Deep Neural Network for lung and heart dose prediction in breast treatments
PH-0287

Abstract

Transfer learning and Deep Neural Network for lung and heart dose prediction in breast treatments
Authors: Perez-Alija|, Jaime(1)*[jperezalija@santpau.cat];Gallego|, Pedro(1);Lizondo|, Maria(1);Nuria|, Jornet(1);Latorre-Musoll|, Artur(1);Valverde-Pascual|, Itziar(1);Barceló-Pagès|, Marta(1);Garcia-Apellaniz|, Nagore(1);Carrasco de Fez|, Pablo(1);Delgado-Tapia|, Paula(1);Simon Garcia|, Pablo(1);Adria Mora|, Mar(1);Ruiz Martinez|, Agustí(1);Ribas|, Montse(1);Ambroa|, Eva(2);
(1)Hospital de la Santa Creu i Sant Pau, Medical Physics, Barcelona, Spain;(2)Consorci Sanitari de Terrassa, Medical Physics Unit- Radiation Oncology Department, Terrassa, Spain;
Show Affiliations
Purpose or Objective

Generating a convolutional neural network (CNN) model to predict lung and heart dose-volume histograms (DVH) in breast cancer patients with lymph nodes treated with 3D-CRT would help in the technique decision process. Usually, the work done in dose prediction using CNNs does not consider the plan quality of the training data. To ensure this quality, we propose a method for outliers detection within the dataset that can be used as a DVH predictor.

Material and Methods

We selected 195 patients with left breast cancer treated with 3D-CRT. We included patients with axillary and supraclavicular lymph nodes but excluded those with an internal mammary nodal (IMN) chain.

For the model creation, we trained the CNN renormalizing all plans to 2 Gy/fraction, to take into account different prescribed doses. For our CNN model, we implemented a transfer learning approach using a pre-trained VGG-16 and replacing its three last layers with a fully connected neural network.

Input data was the planning CT contour information. Output was a 2D lung and heart DVH for every slice.  All slices were subsequently added up to account for the final whole OAR DVH.

For the outliers detection, we partitioned our set in training, validation, and test (176, 10, and 10 patients, respectively). First, we trained the CNN with early stopping. Second, we evaluated how good our model fitted the data in the test set and searched for the presence of any potential outlier using the sum of residuals method to measure the discrepancy between the predicted and the clinical approved DVH; we defined an outlier as any prediction having a sum of residuals greater than one standard deviation from the population mean value. Finally, we repeated this two-step process using different partitions, until all the patients contained in the first training set were once in the test set. At every iteration, we initialized all the CNN parameters to avoid information bleeding.

Once we selected all potential outliers, one researcher (M.L.) proceeded to re-optimized all the plans. We recalculated the sum of residuals for them and elaborated a confusion matrix with the model results.
Results

Our CNN model detected a total of 23 out of 195 patients as having a suboptimal plan. After the reoptimization step, all patients but one were not considered as an outlier anymore. Our false-positive result was of a patient with an IMN chain. We consequently excluded this patient from our dataset since having an IMN was an exclusion criterion.

Figure I shows the clinically approved, replanned, and predicted, with confidence intervals,  lung and heart DVHs for three patients where the model was applied.

Figure 1
Conclusion

We validated our CNN model as a reliable method to account for outliers within a dataset.

We developed an accurate model for DVH prediction in breast cancer patients. This work will allow us to discriminate beforehand which patients will not fulfill dose constraints with 3DCRT and  would benefit from other techniques.