Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Automated breast cancer detection using deep learning based object detection models have achieved high sensitivity, but often struggles with high false positive rate. While radiologists possess the ability to analyze and identify malignant masses in mammograms using multiple views, it poses a challenge for deep learning based models. Inspired by how object appearance behave across multiple views in natural images, researchers have proposed several techniques to exploit geometric correspondence between location of a tumor in multiple views and reduce false positives. We question clinical relevance of such cues. We show that there is inherent ambiguity in geometric correspondence between the two mammography views, because of which accurate geometric alignment is not possible. Instead, we propose to match morphological cues between the two views. Harnessing recent advances for object detection approaches in computer vision, we adapt a state-of-the-art transformer architecture to use proposed morphological cues. We claim that proposed cues are more agreeable with a clinician’s approach compared to the geometrical alignment. Using our approach, we show a significant improvement of 5% in sensitivity at 0.3 False Positives per Image (FPI) on benchmark INBreast dataset. We also report an improvement of 2% and 1% on in-house and benchmark DDSM dataset respectively. Realizing lack of open source code base in this area impeding reproducible research, we are publicly releasing source code and pretrained models for this work.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1306_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1306_supp.pdf

Link to the Code Repository

https://mammo-iitd-aiims.github.io/CEN

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Jai_Follow_MICCAI2024,
        author = { Jain, Kshitiz and Rangarajan, Krithika and Arora, Chetan},
        title = { { Follow the Radiologist: Clinically Relevant Multi-View Cues for Breast Cancer Detection from Mammograms } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15001},
        month = {October},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a framework for breast cancer detection from mammograms by integrating morphological cues from regions of interest (ROIs) in complementary views, the two mammography views MLO and CC, mimicking the radiologists’ practices. This innovative approach allows for interpretable ROIs without relying on problematic geometric alignment methods. The model demonstrates significant advancements in detection accuracy, achieving improved sensitivity levels compared to the state-of-the-art. Additionally, the framework is designed to be adaptable to future deep learning models, promoting transparency and reproducibility in breast cancer detection research.

  • Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The strengths of the paper include: 1- The paper introduces a framework for breast cancer detection that focuses on integrating morphological cues from ROIs in complementary views, deviating from traditional geometric alignment methods. This approach challenges the typical perception in multi-view breast cancer detection, offering a fresh perspective on leveraging visual features for improved precision. 2- The proposed framework is designed to be easily integrated with various deep learning detection models, showcasing its flexibility and readiness for future advancements in the field. This adaptability enhances the scalability and applicability of the approach, ensuring its relevance in evolving research landscapes. 3- The authors release the source code and pre-trained models to promote transparency and reproducibility in breast cancer detection research.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    The weaknesses of the paper include: 1- While the paper emphasizes the use of morphological features from ROIs for breast cancer detection, it lacks detailed insights into the specific morphological features utilized, the extraction process, and the rationale behind their selection. The authors should show some figures for the visualization of the DL models used using interpretability methods such as GRADCAM, LIME or RISE, etc… Providing more clarity on the morphological feature extraction methodology would enhance the reproducibility and understanding of the proposed approach. 2- The paper does not extensively discuss the computational complexity of the proposed framework, including inference time, resource requirements, and scalability. Understanding the computational overhead associated with implementing the method is crucial for assessing its practical feasibility in real-world clinical settings. A more thorough analysis of computational aspects would provide valuable insights for potential adopters of the framework. For instance, instead of mentioning the venue of the papers used in the comparison, the authors should mention the number of the trained parameters of each model and the FLOPs number. 3- The paper could benefit from a more in-depth discussion of the limitations of the proposed framework. Addressing potential challenges, constraints, or scenarios where the approach may not perform optimally would provide a more balanced perspective on the applicability and scope of the method. Acknowledging limitations is essential for guiding future research directions and refining the proposed framework.

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The submission has provided an anonymized link to the source code, and their dependencies. The paper has strengths in terms of reproducibility, such as the commitment to sharing code and models, there could be additional steps in the supplementary materials taken to enhance reproducibility further, such as providing detailed documentation, ensuring data availability, and facilitating easy access to supplementary materials. By addressing these aspects, the authors can strengthen the reproducibility of their work and promote greater transparency and trust in the research outcomes.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    • The paper introduces a novel framework for breast cancer detection that leverages multi-view cues from mammograms. The integration of morphological features from ROIs in the other view is a significant methodological innovation. To further enhance the methodological contribution, consider discussing the computational complexity, and detailing the morphological feature extraction process.
    • The clinical relevance of the proposed framework is evident in its focus on interpretable ROIs without relying on problematic geometric alignment. This approach aligns well with the clinical needs for effective breast cancer detection. To strengthen the clinical applicability, consider conducting additional validation studies on diverse datasets beyond INBREAST and DDSM datasets to demonstrate the generalizability of the framework in real-world clinical settings, such as Nguyen, H.T., Nguyen, H.Q., Pham, H.H. et al. VinDr-Mammo https://doi.org/10.1038/s41597-023-02100-7, and RSNA https://www.rsna.org/rsnai/ai-image-challenge/screening-mammography-breast-cancer-detection-ai-challenge
    • To facilitate clinical translation, consider providing detailed documentation, ensuring data availability, and including supplementary materials with additional experimental results and analysis scripts to support reproducibility and transparency, especially for morphological cues.
    • Consider discussing how the proposed framework may impact underserved populations or diverse patient groups. To promote health equity, consider evaluating the performance of the framework across diverse demographic groups and discussing potential implications for improving access to early detection and treatment.
    • Whether Cosine Similarity alone is sufficient as a loss function depends on the specific objectives of the task and the nature of the data being analyzed. While Cosine Similarity can capture the directional similarity between vectors, it may not account for other important factors such as magnitude differences, scale variations, or specific characteristics of the data distribution. I think the authors could use the Contrastive loss that is commonly used for learning embeddings by pulling together embeddings of similar instances and pushing apart embeddings of dissimilar instances in the feature space. It encourages similar instances to be close to each other and dissimilar instances to be far apart.
    • The authors should apply an ablation study showing the effect of each component on the proposed framework. For instance use the effect of using ResNet50 or others on the features extraction, as well the authors should compare the results of the proposed MLP to for example, Fully connected layers (FC).
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Mainly, the paper requires more experimental results. For example: 1- Exploring a range of loss functions beyond Cosine Similarity and backbones beyond REsNet50, can provide valuable insights into the effectiveness of different optimization strategies for minimizing dissimilarity between features in the breast cancer detection framework. By conducting comparative experiments and evaluating the impact of alternative loss functions and backbones on model performance, the authors can make informed decisions on the most suitable approach for enhancing the interpretability and accuracy of the breast cancer detection system. 2- The breast cancer detection task involves intricate relationships between features extracted from different views. Changing the model structures can help the model capture complex feature interactions, and account for variations in feature distributions. 3- The paper lacks detailed insights into the specific morphological features utilized, in the extraction process

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper describes a novel false positive reduction strategy for detection of lesions in multi-view breast mammography acquisitions. The idea is to leverage morphological cues, emulating the approach of expert readers, instead of relying on ambiguous geometric features. The results are compelling, with the authors promising to release both source code and pre-trained models.

  • Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Novel formulation of a false-positive reduction strategy leveraging morphological cues - called context embedding network.
    • Flexibility of the method: any candidate generation model can be used. Training the embedding network seems straightforward.
    • The paper is generally well written, and easy to follow.
    • The evaluation is extensive, based on multiple datasets (in-house and public), with comparison against multiple solutions.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    • I would recommend rephrasing the motivation section in the introduction to a “gentler” reference to the limitation of related work. The statement “geometric alignment is irrelevant” is harsh; perhaps “ambiguous”, “challenging to exploit”, “limited” … but not “irrelevant”. Rephrasing the introduction/motivation in a neutral way, I believe will further strengthen the message of the paper.
    • A thorough revision of the mathematical notation throughout the paper is needed: e.g., the variable m is used as index to refer to MLO, as variable i=1…m, and as MCS score m_{ij}.
    • I highly recommend a revision of the main message of the paper, contained also in the title: “breast cancer detection”. Throughout the paper “cancer”, “malignancy” and “lesion” are interchangeable used. Is really cancer, as disease, being detected? I would argue that lesions are being detected in such mammography images; while the malignancy of these lesions and actual cancer diagnosis are another question that is not addressed by this paper. It’s very important to be super-precise.

    Further questions, small improvements:

    • Abstract: “behave” -> “behaves”
    • Abstract: “question clinical” -> “question the clinical”
    • “Distance of cancer from the nipple area”: please see my comment above on cancer, being a disease.
    • “to split the DDSM dataset to split”
    • Fig 3: Please use consistent color for each method across the graphs.
    • Fig 4 is quite confusing to me, it does not seem consistent which results are shown in the “first pair vs. second pair”. I would recommend a simplification, perhaps reducing the number of cases and discussing shortly what is observed in each case.
    • Fig 1, supplementary material: any clinical reasoning around the varying distribution?
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Do you have any additional comments regarding the paper’s reproducibility?

    Great!

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    Please see my answers under points 5 and 6. I have included there all my comments and elaborated on the changes that in my view would further improve the paper.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The revision of the introduction/motivation and the revision of the mathematical notation are important for acceptance.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #3

  • Please describe the contribution of the paper

    The main argument of this work is that for a flexible organ such as the breast, the geometric alignment cues used in multi-view work are irrelevant. So the authors use a novel framework to model the relationship between two views in mammography. The framework greatly improves the accuracy of SOTA detection in the DDSM dataset and is validated on an in-house dataset.

  • Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The authors use a novel framework to model the relationship between two views in mammography. The data are adequate and the experiment is complete.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    The rate clarity of the presentation of the methodology section needs to be improved. Specific changes have been noted in the detailed comments.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. Please introduce the full name of the MCS score when it was first introduced.
    2. In “Context Embedding Network (CEN)”, when initializing the ResNet50 weights, the authors pre-trained the model by classifying the predicted bounding boxes in the training set. Is the benign/malignant distinction based on the confidence of malignant tumors in the predicted box? Why is it necessary to pre-train in this way, please explain in detail.
    3. In “Context Embedding Network (CEN)”, the authors concatenate the feature embedding from the ResNet50 model with the bounding box information. Is the bounding box information the location of the bounding box?
    4. What exactly is the process of “Refinement” in Fig. 2? Please state it clearly in the Methodology section.
    5. References 35 and 36 are duplicated.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The article’s experiments are relatively complete. The code is promised to be reproducible. However, the description of the methodology section is not clear enough and some details need to be added.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Author Feedback

We thank the reviewers for their insightful comments and are grateful for the opportunity to improve our manuscript based on their feedback. We observe an inference time of 500ms for a pair of mammogram images (MLO and CC) on a V100 GPU, requiring less than 8GB of GPU memory. We will release our code to ensure easy reproducibility of our results. Our methodology, currently designed for one breast, can be extended to bilateral studies, making it more applicable to clinical settings. Additionally, we will update and maintain color consistency in Figure 3, correct all typos, eliminate duplication in references, and improve mathematical notation in the camera-ready version of the manuscript. Thank you once again for your valuable feedback.



Meta-Review

Meta-review not available, early accepted paper.



back to top