Improved Genotype-Guided Deep Radiomics Signatures for Recurrence Prediction of Non-Small Cell Lung Cancer

Deep Features from Pretrained Networks Do Not Outperform Hand-Crafted Features in Radiomics

In radiomics, utilizing features extracted from pretrained deep networks could result in models with a higher predictive performance than those relying on hand-crafted features. This study compared the predictive performance of models trained with either deep features, hand-crafted features, or a combination of these features in terms of the area under the receiver-operating characteristic curve (AUC) and other metrics. We trained models on ten radiological datasets using five feature selection methods and three classifiers. Our results indicate that models based on deep features did not show an improved AUC compared to those utilizing hand-crafted features (deep: AUC 0.775, hand-crafted: AUC 0.789; p = 0.28). Including morphological features alongside deep features led to overall improvements in prediction performance for all models (+0.02 gain in AUC; p < 0.001); however, the best model did not benefit from this (+0.003 gain in AUC; p = 0.57). Using all hand-crafted features in addition to the deep features resulted in a further overall improvement (+0.034 in AUC; p < 0.001), but only a minor improvement could be observed for the best model (deep: AUC 0.798, hand-crafted: AUC 0.789; p = 0.92). Furthermore, our results show that models based on deep features extracted from networks pretrained on medical data have no advantage in predictive performance over models relying on features extracted from networks pretrained on ImageNet data. Our study contributes a benchmarking analysis of models trained on hand-crafted and deep features from pretrained networks across multiple datasets. It also provides a comprehensive understanding of their applicability and limitations in radiomics. Our study shows, in conclusion, that models based on features extracted from pretrained deep networks do not outperform models trained on hand-crafted ones.

Siamese Delta Network with Multimodality Fusion for Systemic Treatment Response Prediction in Pancreatic Cancer

Systemic treatment is a main way for pancreas cancer patients that are ineligible for surgery. A subgroup of patients showed good response to systemic treatment and the rest received limited benefits. CT images provide a non-invasive way to assess the treatment response. Alternative non-image methods include radiology analysis, tumor marker analysis and combination analysis. To combine the image and non-image data, we propose the Siamese Delta Network with Multimodality Fusion (SDN-MF) to predict systemic treatment response in an end-to-end way. First, a Siamese Delta Network (SDN) is designed to process pre-treatment and pre-surgery CT images and get the image feature changes to predict response. Then, patients' characteristics from EMR and alternative analysis results forms non-image data, which is incorporated into SDN with a multimodality fusion (MF) module. The proposed SDN-MF is evaluated on a private dataset and achieves average AUC value of 0.883 with five cross-validation. Comparison among image-only, non-image-only, and fusion models verifies the superior of multimodality model in predicting systemic treatment response of pancreas cancer patients.

A CT-based transfer learning approach to predict NSCLC recurrence: The added-value of peritumoral region

Non-small cell lung cancer (NSCLC) represents 85% of all new lung cancer diagnoses and presents a high recurrence rate after surgery. Thus, an accurate prediction of recurrence risk in NSCLC patients at diagnosis could be essential to designate risk patients to more aggressive medical treatments. In this manuscript, we apply a transfer learning approach to predict recurrence in NSCLC patients, exploiting only data acquired during its screening phase. Particularly, we used a public radiogenomic dataset of NSCLC patients having a primary tumor CT image and clinical information. Starting from the CT slice containing the tumor with maximum area, we considered three different dilatation sizes to identify three Regions of Interest (ROIs): CROP (without dilation), CROP 10 and CROP 20. Then, from each ROI, we extracted radiomic features by means of different pre-trained CNNs. The latter have been combined with clinical information; thus, we trained a Support Vector Machine classifier to predict the NSCLC recurrence. The classification performances of the devised models were finally evaluated on both the hold-out training and hold-out test sets, in which the original sample has been previously divided. The experimental results showed that the model obtained analyzing CROP 20 images, which are the ROIs containing more peritumoral area, achieved the best performances on both the hold-out training set, with an AUC of 0.73, an Accuracy of 0.61, a Sensitivity of 0.63, and a Specificity of 0.60, and on the hold-out test set, with an AUC value of 0.83, an Accuracy value of 0.79, a Sensitivity value of 0.80, and a Specificity value of 0.78. The proposed model represents a promising procedure for early predicting recurrence risk in NSCLC patients.