How well does your proposed model detect tiny lesions in medical images?

[linhduongtuan]

I really appreciate your work. However, to make it easier to follow, I would suggest cleaning up your repository and improving the README file. This would make it worthwhile for others to reimplement your work.

Additionally, as you know, recognizing lesions in medical modalities is often challenging, especially when they are very small. Therefore, I am curious if your work can detect tiny abnormalities well. It would be helpful if you could provide evidence of this, such as object detection and segmentation of breast ultrasound data for cancer classification.

Linh

[duyhominhnguyen]

Hi @linhduongtuan, thank for your interest to our work and recommendations. Can you please specify which part in the Readme.txt you want to have a better version? We tried to make it simple by creating a Table of Content that you can click and jump directly corresponding section. Anyway, just suggest us your ideas.

Regarding your questions on "object detection and segmentation of breast ultrasound data for cancer classification", indeed, we already provided segmentation results on this setting (BUID dataset) (Table 2 in paper). You can also reproduce results for this case using our Github. We will soon add another setting for retinopathy-related lesions in the FGADR dataset whose segment regions are tiny. In any way, further exploiting LVM-Med performance on your settings are welcome. Thank you.

[linhduongtuan]

@duyhominhnguyen Thank you for providing detailed clarifications. Based on my experiments with the BUID dataset for segmentation tasks, I have observed that overall metrics such as Dice and mIoU tend to yield satisfactory results, primarily due to the data imbalance (with a significantly higher number of benign instances compared to malignant ones). However, when evaluating these metrics for individual categories, they may not perform well for the malignant class. In such cases, relying solely on overall Dice or mIoU scores is akin to emphasizing overall accuracy instead of F1-score in classification problems, especially in medical modalities where imbalanced data is prevalent.

Regarding the README file, I would like to suggest improvements in the folder and file structures to enhance readability and ease of access. Currently, it is necessary to refer to the Table of Contents and click on specific instructions to delve into the details. It would be more beneficial if the file structure were made more intuitive and user-friendly, facilitating easier navigation and comprehension.

[duyhominhnguyen]

@linhduongtuan Thanks for your information. Basically, we just evaluate several classes followed by prior works on average. In any way, further analyzing LVM-Med performance and suggesting efficient ways to fine-tune specific cases as desired are welcome! We are happy to see such improvements from the community.

Regarding the Readme, we will take into account your opinion and will wait for more feedback before updating with a new version.