PaperlayerView exact revision on arXiv

Source-linked AI summary

SD-CNN: a Shallow-Deep CNN for Improved Breast Cancer Diagnosis

Fei Gao, Teresa Wu, Jing Li, Bin Zheng, Lingxiang Ruan, Desheng Shang, Bhavika Patel

arXiv:1803.00663v2cs.CV

TL;DR

FFDM screening has limited sensitivity in some groups, while CEDM provides recombined information but is less accessible. The paper proposes SD-CNN to synthesize virtual recombined images from FFDM and use CNN-derived features for diagnosis, reporting improved accuracy when recombined or virtual recombined information is added.

  • Problem

    FFDM has relatively low detection sensitivity in several high-risk groups, while CEDM information is less accessible because CEDM systems are not widely available.

  • Method

    SD-CNN combines a shallow CNN that renders virtual recombined images from FFDM with a deep CNN that extracts diagnostic features from mammographic and recombined images.

  • Results

    Adding recombined or virtual recombined imaging improves breast cancer diagnosis accuracy in the reported experiments.

  • Takeaways & Limitations

    The study supports using synthesized recombined-image information to extend some diagnostic advantages of CEDM to FFDM-based diagnosis.

  • Takeaways & Limitations

    The ResNet feature generator is a black box whose extracted features may be difficult for physicians to interpret.

Abstract

from arXiv · show

Breast cancer is the second leading cause of cancer death among women worldwide. Nevertheless, it is also one of the most treatable malignances if detected early. Screening for breast cancer with digital mammography (DM) has been widely used. However it demonstrates limited sensitivity for women with dense breasts. An emerging technology in the field is contrast-enhanced digital mammography (CEDM), which includes a low energy (LE) image similar to DM, and a recombined image leveraging tumor neoangiogenesis similar to breast magnetic resonance imaging (MRI). CEDM has shown better diagnostic accuracy than DM. While promising, CEDM is not yet widely available across medical centers. In this research, we propose a Shallow-Deep Convolutional Neural Network (SD-CNN) where a shallow CNN is developed to derive "virtual" recombined images from LE images, and a deep CNN is employed to extract novel features from LE, recombined or "virtual" recombined images for ensemble models to classify the cases as benign vs. cancer. To evaluate the validity of our approach, we first develop a deep-CNN using 49 CEDM cases collected from Mayo Clinic to prove the contributions from recombined images for improved breast cancer diagnosis (0.86 in accuracy using LE imaging vs. 0.90 in accuracy using both LE and recombined imaging). We then develop a shallow-CNN using the same 49 CEDM cases to learn the nonlinear mapping from LE to recombined images. Next, we use 69 DM cases collected from the hospital located at Zhejiang University, China to generate "virtual" recombined images. Using DM alone provides 0.91 in accuracy, whereas SD-CNN improves the diagnostic accuracy to 0.95.

1. INTRODUCTION

FFDM has limited detection sensitivity in some high-risk groups, motivating CEDM and the SD-CNN approach. The study uses CNNs both to exploit recombined CEDM information and to synthesize virtual recombined images from FFDM.

  • Motivation: FFDM screening has relatively low detection sensitivity in several high-risk groups, including women with dense breasts.Its overall cancer detection accuracy is reported as 0.75 to 0.85 in the general population, with lower accuracy in some high-risk subgroups.
  • Motivation: CEDM combines mammography with contrast-enhanced imaging to provide information related to tumor neoangiogenesis.The modality uses an iodinated contrast agent and includes low- and high-energy acquisition techniques.
  • Proposed approach: SD-CNN is proposed as a CADx scheme that uses information available from CEDM to improve malignancy diagnosis with FFDM.The approach addresses the difficulty of image interpretation and the limited accessibility of CEDM systems.
  • CNN rationale: Deep CNNs can extract features at different abstraction levels but generally require large labeled datasets and substantial computational resources.These requirements can be difficult to satisfy in medical imaging applications.
  • CNN rationale: CNNs can also learn nonlinear mappings between input and output image patches to generate synthetic medical images.The paper places its shallow-CNN image-rendering strategy in this broader application of CNNs.
  • Study objectives: The study tests whether deep CNNs can exploit recombined CEDM images and whether shallow CNNs can synthesize virtual recombined images from FFDM.These two hypotheses are evaluated using CEDM and FFDM datasets.

2. MATERIALS

The study uses two separate imaging datasets: an institutional CEDM dataset from Mayo Clinic Arizona and the public INbreast FFDM dataset. The materials include clinical CEDM examinations, lesion annotations, and CC/MLO mammographic views.

  • Datasets: Two separate datasets are used: an institutional dataset from Mayo Clinic Arizona and the public INbreast dataset.The datasets support the study’s CEDM and FFDM imaging analyses.
  • Mayo Clinic dataset: The Mayo Clinic dataset contains CEDM examinations acquired between August 2014 and December 2015 under an IRB-approved protocol.Patients had BI-RADS ratings of 4 or 5 on their original FFDM screening images, and CEDM was performed before biopsy.
  • Mayo Clinic dataset: Mayo Clinic CEDM examinations include low-energy and recombined images viewed in CC and MLO projections.Figure 1 illustrates example views for cancer and benign cases.
  • Mayo Clinic dataset: A breast radiologist independently interpreted the Mayo Clinic mammograms and outlined lesion contours on recombined images.The images were de-identified, transferred to a research database, and prepared for blinded reading.
  • INbreast dataset: The INbreast dataset is an online accessible full-field digital mammographic database acquired with a Siemens MammoNovation system.The dataset provides FFDM images for the public-dataset component of the study.
  • INbreast dataset: The study uses both CC and MLO views for feature extraction from the INbreast images.Figure 2 presents CC on the left and MLO on the right, with lesions highlighted by green circles.

3 METHODOLOGY

The SD-CNN combines shallow CNN rendering with deep CNN feature extraction to use real or virtual recombined breast images for diagnosis. ResNet features are organized across layers and passed to boosting-based classifiers.

  • The SD-CNN first learns relationships between low-energy and recombined images, then renders virtual recombined images from FFDM for combined diagnosis.For CEDM, the workflow can begin directly with the deep CNN.
  • Preprocessing: Tumor-region bounding boxes are used instead of whole images because the study focuses on tumor diagnosis.The preprocessing identifies minimum-area boxes containing the tumor region.
  • Shallow-CNN: Virtual Image Rendering: A four-layer shallow CNN models the latent relationship between input and output images for synthetic image rendering.
  • Deep-CNN: Feature Generation: ResNet uses shortcut building blocks to mitigate vanishing gradients, and ResNet-50 is selected to balance computational efficiency and accuracy.ResNet-50 contains 3, 4, 6, and 3 blocks of the four respective block types.
  • Classification: Features extracted from the final layer of each ResNet block type are supplied to gradient boosting trees for classification.The four extraction layers are 10, 22, 40, and 49; boosting reduces bias and variance through an ensemble of weak learners.

4 EXPERIMENTS AND RESULTS

The experiments test whether recombined-image features improve CEDM diagnosis and whether shallow CNNs can generate useful virtual recombined images from FFDM. Across both settings, adding recombined or virtual recombined features improves reported diagnostic performance.

  • Experiment I: 0.89 accuracy and 0.91 AUC were achieved using LE plus recombined images, compared with 0.85 accuracy and 0.84 AUC using LE alone.Sensitivity and specificity also increased from 0.89 and 0.80 to 0.93 and 0.86, respectively.
  • Experiment I: Recombined-image features contributed 23.16% of impurity reduction, while LE features contributed 76.84% across the 99-feature model.The recombined features were associated with improving accuracy from 0.85 to 0.89.
  • Experiment II: A 4-layer shallow CNN learned the nonlinear mapping from LE images to recombined images using 98 images from 49 CEDM cases.The training data comprised 220000 paired 15×15 patches, with 10 validation images and 88 training images.
  • Experiment II: 0.031 MSE was obtained on 10 validation images when comparing virtual recombined images with true recombined images.The standard deviation was 0.021, and sample images restored tumor shape and some details.
  • Experiment II: Virtual recombined images excluded dense tissues surrounding tumors and filtered suspicious masses in benign cases, preserving a reported recombined-image advantage.These images were rendered from 89 FFDM cases using sliding 15×15 input windows and 3×3 output patches.
  • Experiment II: 0.92 ± 0.14 mean AUC was achieved using FFDM plus virtual recombined images, versus 0.87 ± 0.12 using FFDM features alone.The two models had similar sensitivities, while their specificities varied substantially.
  • Experiment II: The virtual-recombined approach achieved an AUC of 0.92 but 0.90 accuracy against a literature method reporting 0.91 AUC and 0.95 accuracy.The authors report outperforming six of eight compared algorithms in both accuracy and AUC, while remaining inferior in accuracy to the best Dhungel et al. approach.

5 DISCUSSION AND CONCLUSION

The study uses SD-CNN to improve breast lesion classification by combining deep-CNN features from recombined CEDM images with shallow-CNN-generated virtual recombined images. Recombined features improved performance for both CEDM and FFDM-based diagnosis, while feature interpretability and patch-size effects remain future work.

  • Accuracy increased from 0.84 to 0.90 and AUC from 0.87 to 0.92 when virtual recombined imaging features were added to FFDM-based classification.A 4-layer shallow CNN learned the nonlinear association between LE and recombined CEDM images to render virtual recombined images from FFDM.
  • The SD-CNN addresses limited CEDM accessibility by using virtual recombined images with FFDM for breast cancer diagnosis.
  • The trained ResNet is a black-box feature generator, making its extracted features difficult for physicians to interpret.
  • Future work will assess clinical interpretations of extracted features and the effects of different input and output patch sizes.
Loading 1803.00663v2…