Pre-Training Autoencoder for Lung Nodule Malignancy Assessment Using CT Images
Round 1
Reviewer 1 Report
In the proposed paper, authors explore the use of Transfer Learning to assess the malignancy of lung nodules using CT images. The authors use a Convolutional Autoencodes and a multi-stage training strategy as a feature extraction method, used as an input to a classifier that outputs the final prediction. Main point of the article is that features from networks learned for other tasks can be used for nodule malignancy prediction, advantages of which are discussed in the text.
The article is generally well written and well structured. Only minor changes are required to make some sentences correct and a bit more understandable. The problem motivation is clearly outlined with an adequate overview of the state of the art, both in classical approaches and TL approaches. The aricle addresses some of the challenges in the field, such as data labeling and interrates inconsistencies. The three contributions that authors underline, lower computational resource consumption, less need for annotated data and easier feature extraction are all important aspects in today's medical image processing. The presented methodology is sound and research objectives relevant to the field. Limitations of the study are properly addressed and discussed in details, with some of the future improvements mentioned in the Discussion section.
-
Comments:
- Intrarater variability over classes could be reported to better illustrate this phenomena. Can you discuss interrater variability generally a bit more, as it gives a good perspective of human-level performance?
- Is it possible to use other classifiers, with the exctracted features as inputs? Did you consider any?
- Is is possible to visualize some of the layers of the encoder? It would be interesting to know and discuss which image features are important for malignancy prediction.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The authors followed the research methodology related to the lung module malignancy classification by TL approach, based on CAE with LIDC-IDRI dataset. It would be interesting that next time, top perform on their own dataset, as an important contribution to extend the existing dataset. The AUC is about 0.936 accuracy, but in the literature is mentioned better results (references 10 and 11) considering other methods. A justification of the conducted procedure with a less accuracy would be useful and further improvement steps.
Author Response
"Please see the attachment.
Author Response File: Author Response.pdf