Purpose: To mimic dose of carcinoma cervix treatment plans using a novel convolutional deep neural network (cDNN) architecture.
Methods: Retrospectively, a total of 235 patients of cervix carcinoma (Stage I-IV) from the year 2015-2021 who had received radiotherapy treatment using VMAT Techniques (@6MV X-rays, 2 full Arcs, collimator angle: 150 and 3450, prescription: 46Gy@23Fx). Each patient had 5 contours: PTV, Bladder, Left femoral head, Right femoral head, and Rectum. Data is divided into 3 sets: The training dataset is having 201 patients, Validation dataset is having 19 patients and the testing dataset is having 15 patients. The proposed model architecture consists of multiple convolutional layers followed by batch Normalization and a Leaky rectified linear unit. The output of each layer is considered with its corresponding feature map followed by spatial dropout and ReLu. For finer results, the output is again convolved through same layers. The training of the model was done in 190 epochs.
Results: Dose Volume Histogram score and Dose score is 1.85 and 2.015 respectively. The average percentage relative change in maximum and mean doses of all structures is within 2.98% and 1.03% of the prescription dose. The average Dice similarity coefficient (DSC) score is 0.94 when comparing the predicted vs. true isodose volumes between 0% and 90% of the prescription dose. The gamma index with criteria 3%/3 mm and 3%/2 mm has an average passing rate of 99.18 and 97.46 respectively. The Homogeneity and Paddick's Conformity Index is 0.067 and 0.76. The trained model is able to predict the dose distribution accurately for test patients within a clinically acceptable range.
Conclusion: The proposed cDNN architecture has the potential to be utilized in radiotherapy treatment planning as a guiding tool in the radiation oncology department.
DICOM-RT, Treatment Planning, Warping
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation