Jul 31, 2021
This research paper aims to develop a deep-learning Convolutional Neural Network (CNN) for predicting the complete pathological response (pCR) in oesophageal cancers after Neo-Adjuvant ChemoRadioTherapy (NACRT) based on pre-NACRT imaging alone. The study utilised a dataset of 211 oesophageal cancer patients who underwent NACRT followed by surgery, with 192 patients used for model development and validation.
The deep learning methodology involved image pre-processing and augmentation, where tumours were delineated on pre-NACRT CT imaging and converted to NIFTI-2 format. A customised VGG-16 architecture was used to handle 3D NIFTI-2 data. Hyperparameter optimisation was performed using random search, Bayesian optimisation, and grid search techniques.
The results showed that the CNN architecture comprised nine 3D convolutional layers and achieved a mean accuracy of 0.75 ± 0.02 in the training set and an accuracy of 0.74 in the validation set. The F1-score for predicting pCR was 0.8, and the model's overall accuracy was 0.68. Activation maps of the final convolutional layer revealed that the region within the Gross Tumor Volume (GTV), excluding the oesophagus lumen, was most predictive of pCR.
In conclusion, the study demonstrates that deep learning methods can predict pCR after NACRT in oesophageal cancers with an accuracy of 68%. However, external validation is necessary before clinical adoption.
