Domain Adaptation for Deviating Acquisition Protocols in CNN-Based Lesion Classification on Diffusion-Weighted MR Images

Jennifer Kamphenkel, Paul F. Jäger, Sebastian Bickelhaupt, Frederik Bernd Laun, Wolfgang Lederer, Heidi Daniel, Tristan Anselm Kuder, Stefan Delorme, Heinz-Peter Schlemmer, Franziska König, Klaus H. Maier-Hein

TL;DR

This paper tackles the challenge of applying deep learning models for breast cancer classification on MRI scans when different hospitals use varying imaging protocols. They introduce a "model-based domain adaptation" technique that uses a physics-based model (DKI) to reconstruct missing or altered MRI data, allowing a pre-trained deep learning model to work effectively without retraining. This approach significantly improves classification performance in real-world scenarios with diverse imaging setups.

What Problem Does This Paper Address?

Deep learning, specifically Convolutional Neural Networks (CNNs), has shown great promise in automatically classifying breast cancer lesions from diffusion-weighted MRI (DWI). However, a major hurdle for widespread clinical adoption is the "input dependence" of these CNNs. This means a CNN trained on DWI data acquired with a specific protocol (e.g., certain b-values or gradient strengths) may perform poorly when presented with data from a different protocol. This is a critical issue because hospitals worldwide use diverse MRI scanners and protocols, making it impractical to retrain models for each new site. The paper aims to bridge this gap by developing a method that makes CNNs robust to these variations in acquisition protocols, enabling seamless deployment across different clinical settings.

Methods

The core innovation is Model-Based Domain Adaptation (MBDA). Instead of directly feeding altered input data into the CNN, MBDA first uses a biophysical model, Diffusion Kurtosis Imaging (DKI), to estimate key diffusion properties (like ADC and AKC) from the available, potentially altered, DWI data. The DKI model is then used to synthesize or restore the original DWI input channels that the CNN was originally trained on. This restored data is then fed into the pre-trained CNN for classification.

This approach is contrasted with two other methods:

  • End-to-End (E2E) with Altered Input: The standard CNN approach where altered input data is fed directly into the model without any adaptation. This serves as a baseline for poor performance.
  • Fit-to-End (F2E): This method trains a CNN on the parameters derived from DKI (ADC and AKC) rather than raw DWI images. While it offers some robustness, it relies on the quality of the DKI fit and can be problematic when the DKI model is under-constrained (e.g., with fewer b-values).

The MBDA method essentially uses the DKI model as a "translator" to make the altered input data understandable to the original E2E CNN.

Datasets

The study utilizes a dataset of 221 patients previously used for E2E training. For each patient, DWI images were acquired at four b-values: 0, 100, 750, and 1500 s/mm². The dataset includes 121 malignant and 100 benign lesions, with biopsy results serving as the ground truth. The paper simulates "shifted" and "missing" b-value scenarios to test the robustness of their method against protocol deviations.

Key Insights and Ideas

  • Physics-informed deep learning: The paper elegantly combines the data-driven power of CNNs with the interpretability and physical grounding of DKI. This hybrid approach leverages the strengths of both.
  • Restoring lost information: MBDA doesn't just ignore missing data; it actively reconstructs it using a biophysical model. This is crucial because even a few missing b-values can significantly impact the information available for classification.
  • Avoiding retraining: The primary benefit is enabling the deployment of pre-trained models without the need for costly and time-consuming retraining at every new clinical site. This is a major step towards practical, large-scale AI in medical imaging.
  • Implicit vs. Explicit Domain Adaptation: The paper highlights that while F2E offers implicit domain adaptation by using DKI parameters, MBDA provides a more explicit and effective adaptation by restoring the original input space for the CNN.

Key Results

  • Significant performance improvement: MBDA significantly boosted classification performance in both "shifted" (12.4% increase) and "missing" (16.8% increase) b-value scenarios compared to using altered input directly with E2E.
  • Superiority over F2E: MBDA outperformed F2E (implicit domain adaptation) by a considerable margin, achieving 5.3% higher AUC in the shifted scenario and 12.4% higher AUC in the missing scenario.
  • Robustness to input variations: The method demonstrated a strong ability to handle deviations in acquisition protocols, making CNNs more "agnostic" to specific scanner settings.
  • General redundancy of information: The study noted that even with subsets of three b-values, performance was only moderately decreased compared to using all four, suggesting that the information is somewhat distributed across b-values.

Limitations and Future Work

  • Limited b-value alterations: The study focused on scenarios with one altered or missing b-value, as exploring multiple alterations with the limited number of b-values available was not feasible.
  • No alteration of b=0: The b=0 image (representing proton density) was assumed to be always available, as it's standard practice in most protocols.
  • Extrapolation to large b-values: Performance drops were observed when extrapolating to very large b-values, indicating this remains a challenging area.
  • Future directions: The authors suggest exploring multiple input alterations, applying the method to unsegmented DWI data, investigating generalization on larger datasets, and extending it to other medical entities.

Summary updated

Abstract

End-to-end deep learning improves breast cancer classification on diffusion-weighted MR images (DWI) using a convolutional neural network (CNN) architecture. A limitation of CNN as opposed to previous model-based approaches is the dependence on specific DWI input channels used during training. However, in the context of large-scale application, methods agnostic towards heterogeneous inputs are desirable, due to the high deviation of scanning protocols between clinical sites. We propose model-based domain adaptation to overcome input dependencies and avoid re-training of networks at clinical sites by restoring training inputs from altered input channels given during deployment. We demonstrate the method’s significant increase in classification performance and superiority over implicit domain adaptation provided by training-schemes operating on model-parameters instead of raw DWI images.
BibTeX
@inproceedings{kamphenkel2018domain,
  author = {Jennifer Kamphenkel and Paul F. Jäger and Sebastian Bickelhaupt and Frederik Bernd Laun and Wolfgang Lederer and Heidi Daniel and Tristan Anselm Kuder and Stefan Delorme and Heinz-Peter Schlemmer and Franziska König and Klaus H. Maier-Hein},
  title = {{Domain Adaptation for Deviating Acquisition Protocols in CNN-Based Lesion Classification on Diffusion-Weighted MR Images}},
  booktitle = {RAMBO+BIA+TIA @ MICCAI},
  year = {2018},
  doi = {10.1007/978-3-030-00946-5\_8},
}
Figure