Implementation Strategy of a CNN Model Affects the Performance of CT Assessment of EGFR Mutation Status in Lung Cancer Patients

Abstract

To compare CNN models implemented using different strategies in the CT assessment of EGFR mutation status in patients with lung adenocarcinoma. 1,010 consecutive lung adenocarcinoma patients with known EGFR mutation status were randomly divided into a training set (n = 810) and a testing set (n = 200). The CNN models were constructed based on ResNet-101 architecture but implemented using different strategies: dimension filters (2D/3D), input sizes (small/middle/large and their fusion), slicing methods (transverse plane only and arbitrary multi-view planes), and training approaches (from scratch and finetuning a pre-trained CNN). The performance of the CNN models was compared using AUC. The fusion approach yielded consistently better performance than other input sizes, although the effect often did not reach statistical significance. Multi-view slicing was significantly superior to the transverse method when fine-tuning a pre-trained 2D CNN but not a CNN trained from scratch. The 3D CNN was significantly better than the 2D transverse plane method but only marginally better than the multi-view slicing method when trained from scratch. The highest performance (AUC = 0.838) was achieved for the fine-tuned 2D CNN model when built using the fusion input size and multi-view slicing method. The assessment of EGFR mutation status in patients is more accurate when CNN models use more spatial information and are finetuned by transfer learning. Our finding of the implementation strategy of a CNN model could be a guide to other medical 3D images applications. Compared with other published studies which used medical images to identify EGFR mutation status, our CNN model achieved the best performance in the biggest patient cohort.INDEX TERMS CNN, EGFR, implementation strategy.

A review of transfer learning for medical image classification

“…This is a particularly good practice of science because solutions are shared online for the sake of reproducibility with the dataset. Popular datasets are listed in alphabetical order: BRATS is a dataset that provides The search string applied in Web of Science database was as follows: TS=("CNN" OR "convolutional") AND TS=("medical imag*" OR "clinical imag*" OR "biomedical imag*") AND TS=("transfer learning" OR "pre-trained" OR "pretrained") NOT TS=("novel" OR "propose") Alimentary system Feature extractor [34,35] Fine-tuning scratch [36,37] Bones Feature extractor [38] Genital systems Fine-tuning scratch [39] Nervous system Many [40] Respiratory system Feature extractor [41] Feature extractor hybrid [42] Fine-tuning scratch [43][44][45] Many [46,47] Sense organs Feature extractor [48] Thoracic cavity Feature extractor [49] Endoscopy Alimentary system Feature extractor [50,51] Fine-tuning scratch [52][53][54] Many [55] Mammographic Integumentary system Feature extractor [2]…”

Multi-channel multi-task deep learning for predicting EGFR and KRAS mutations of non-small cell lung cancer on CT images

“…This approach is a non-invasive and easy-toimplement deep learning method. Other studies (27)(28)(29) also show that deep learning models can identify gene mutations in lung cancer.…”

Towards Machine Learning-Aided Lung Cancer Clinical Routines: Approaches and Open Challenges

“…These approaches were based on radiological qualitative features [ 71 , 188 , 189 ]. On the other hand, the features from the CT images can be objective and automatically extracted, such as radiomic or high-level deep features [ 190 , 191 , 192 ]. Additionally, both types of features (semantic features and the automatically extracted) can be used together by the learning models [ 50 ].…”