Project Proposal Feedback

[pooyanjamshidi]

Feedback on Skin Cancer Detection Proposal:

Your project tackles an important healthcare problem, and I appreciate how clearly you've structured your methodology. Skin cancer detection using non-dermoscopic images from the ISIC SLICE-3D dataset addresses a real need in making early skin cancer screening more accessible to populations that might lack access to high-quality medical imaging. Below are my thoughts and suggestions to help you further refine the proposal:

Strengths:

• Well-Defined Problem: You’ve identified a real-world challenge: lack of access to high-quality dermoscopy equipment for early skin cancer detection. Your project aims to solve this by enabling effective detection with lower-quality images, making the solution more accessible to the general public. This is a crucial application with strong potential societal impact.
• Choice of Methodology: Using EfficientNet models to work with non-dermoscopic images is a solid approach. EfficientNet has shown excellent results in image-based tasks while being computationally efficient, which is important since your goal is to enable smartphone-based diagnoses.
• Use of the ISIC SLICE-3D Dataset: The dataset is a great choice, especially since it includes diverse data from multiple institutions across different regions. This should help ensure that your model is generalizable and not overfitted to one specific dataset.
• Clinically-Relevant Metrics: Focusing on sensitivity (True Positive Rate) with an 80% threshold is a good move. In a medical context, particularly with cancer detection, high sensitivity is more important than minimizing false positives since false negatives are far more costly. Your focus on the ROC curve and sensitivity-specific metrics is appropriate for this use case.

Suggestions:

1. Clarify Image Quality Considerations:

   • While you mention that your dataset consists of non-dermoscopic images that could be similar in quality to smartphone images, it would be good to elaborate on how you will handle variability in image quality. Since real-world conditions can produce even lower-quality images (e.g., poor lighting, blurriness), are there plans to include preprocessing techniques like contrast enhancement, noise reduction, or image augmentation? These techniques could help your model become more robust to suboptimal input conditions.

2. Threshold Justification:

   • You’ve set the sensitivity threshold at 80% based on the clinical relevance. It would help to elaborate on why you chose this number specifically. Was it based on established clinical guidelines for skin cancer screening, or is it derived from other models in the literature? This could help ground your choice and lend more support to this threshold.

3. Balancing Sensitivity and Specificity:

   • You mention that false positives can be managed through follow-up procedures, but it would be useful to discuss how you will ensure that specificity is still considered in your evaluation. Even though sensitivity is the priority, very high false positive rates could lead to unnecessary anxiety and over-screening. Perhaps discussing ways to fine-tune the trade-off between sensitivity and specificity could strengthen the clinical applicability of the model.

4. Ensemble Methods and Generalization:

   • Ensemble methods can significantly boost performance, but they can also introduce higher computational overhead. It would be useful to explain how the ensemble of EfficientNet models improves performance in this case. Are you combining predictions to smooth out errors, or are you focusing on improving sensitivity for rare cases? Additionally, what steps will you take to ensure that the ensemble doesn’t overfit the training data? Consider cross-validation strategies and regularization techniques to enhance model generalization.

5. Addressing Class Imbalance:

   • Skin cancer datasets often suffer from class imbalance, where benign lesions far outnumber malignant ones. How do you plan to address this in your model? Techniques such as data augmentation, oversampling, or cost-sensitive learning could be critical to ensuring that your model performs well across both benign and malignant cases.

6. Real-Time Application and Deployment:

   • Since your goal is to create a system that users can run on their smartphones, are there any plans for real-time processing or edge computing strategies? While the model will initially be cloud-based, it might be useful to discuss how you plan to transition to real-time deployment. Could you use model compression techniques, like quantization or pruning, to make the model more lightweight for mobile devices?

7. Privacy and Data Security:

   • You mention that user data will be stored and could be integrated into future training data to improve the model. Make sure to address the privacy and security concerns around storing sensitive medical images. How will you anonymize the data and ensure compliance with data protection laws (e.g., HIPAA or GDPR)?

8. Comparison with Existing Systems:

   • It’s great that you plan to compare your model’s performance against submissions in the ISIC 2024 challenge and the research paper you referenced. Be sure to outline the criteria for these comparisons clearly. Will you compare using the same metrics (AUC, sensitivity, etc.), and what specific aspects of your model do you expect to outperform (e.g., performance on lower-quality images)?

Additional Suggestions:

• Explain the Preprocessing Pipeline in Detail:

  • Since the quality of input data is a significant concern, it would be helpful to detail the image preprocessing pipeline. Will you apply techniques like resizing, normalization, or augmentations (e.g., rotations, flips)? Also, how will you handle the different image resolutions and potential artifacts in real-world smartphone images?

• User Interaction and System Design:

  • Consider expanding a bit more on the system component you mention. How will users interact with the system? Will they upload images via an app, and how will they receive feedback on their results? A clearer outline of the end-to-end user experience could make the project vision even more concrete.

---

Conclusion:

Your proposal addresses a critical healthcare need, and your approach using EfficientNets and ensemble methods is well-suited for the task. Focusing on accessible, non-dermoscopic images while maintaining high sensitivity is the right move for a project aimed at early detection in less specialized settings. As you proceed, make sure to clarify the rationale behind key choices (e.g., sensitivity threshold, preprocessing strategies) and consider the practical aspects of deployment and privacy. This is a well-thought-out project with strong potential—looking forward to seeing how it develops!