[Paper Review] Adaptation of a deep learning malignancy model from full-field digital mammography to digital breast tomosynthesis

Sadanand Singh, Thomas P. Matthews+7 more|arXiv (Cornell University)|Jan 23, 2020

AI in cancer detection20 references4 citations

TL;DR

This paper proposes a domain adaptation method to transfer a deep learning malignancy model trained on full-field digital mammography (FFDM) to digital breast tomosynthesis (DBT) using maximum intensity projection (MIP) images. By applying average histogram matching to reduce domain shift and fine-tuning with SpotTune, the model achieves a ROC AUC of 0.847 on DBT MIP images—significantly outperforming direct testing on MIPs (AUC ~0.75) and demonstrating effective adaptation with limited DBT data.

ABSTRACT

Mammography-based screening has helped reduce the breast cancer mortality rate, but has also been associated with potential harms due to low specificity, leading to unnecessary exams or procedures, and low sensitivity. Digital breast tomosynthesis (DBT) improves on conventional mammography by increasing both sensitivity and specificity and is becoming common in clinical settings. However, deep learning (DL) models have been developed mainly on conventional 2D full-field digital mammography (FFDM) or scanned film images. Due to a lack of large annotated DBT datasets, it is difficult to train a model on DBT from scratch. In this work, we present methods to generalize a model trained on FFDM images to DBT images. In particular, we use average histogram matching (HM) and DL fine-tuning methods to generalize a FFDM model to the 2D maximum intensity projection (MIP) of DBT images. In the proposed approach, the differences between the FFDM and DBT domains are reduced via HM and then the base model, which was trained on abundant FFDM images, is fine-tuned. When evaluating on image patches extracted around identified findings, we are able to achieve similar areas under the receiver operating characteristic curve (ROC AUC) of $\sim 0.9$ for FFDM and $\sim 0.85$ for MIP images, as compared to a ROC AUC of $\sim 0.75$ when tested directly on MIP images.

Motivation & Objective

• To address the challenge of limited labeled DBT data by adapting a pre-trained FFDM malignancy model to DBT exams.
• To reduce domain shift between FFDM and DBT image distributions using image preprocessing techniques.
• To evaluate the effectiveness of transfer learning and adaptive fine-tuning strategies on DBT MIP images with minimal target-domain data.
• To improve malignancy classification performance on DBT without training from scratch on large DBT datasets.

Proposed method

• The study uses maximum intensity projection (MIP) to convert 3D DBT volumes into 2D images for compatibility with existing FFDM models.
• Average histogram matching (HM) is applied to align the intensity distributions of FFDM and DBT MIP images, reducing domain shift.
• A pre-trained deep learning model on FFDM is fine-tuned on DBT MIP images, with and without HM, to adapt to the new domain.
• SpotTune, a learnable fine-tuning policy, is used to selectively update specific residual blocks in the network based on input image characteristics.
• Performance is evaluated using ROC AUC on image patches around radiologically annotated findings.
• The approach leverages transfer learning from a large FFDM dataset to a smaller DBT dataset, minimizing the need for extensive new annotations.

Experimental results

Research questions

• RQ1 Can a deep learning malignancy model trained on FFDM be effectively adapted to DBT MIP images with limited labeled DBT data?
• RQ2 How effective is histogram matching in reducing domain shift between FFDM and DBT image distributions?
• RQ3 Does fine-tuning or SpotTune improve model performance on DBT MIPs compared to direct inference?
• RQ4 How does the performance of the adapted model compare to a baseline model trained directly on DBT MIPs with no pre-training?
• RQ5 What is the relative contribution of preprocessing (HM) versus model adaptation (fine-tuning) in achieving high performance on DBT?

Key findings

• Histogram matching alone achieved the best performance, with a ROC AUC of 0.847 on DBT MIP images, outperforming all other adaptation strategies.
• Without histogram matching, fine-tuning and SpotTune improved performance from a baseline AUC of ~0.75 to 0.837 and 0.830, respectively.
• After histogram matching, the performance gap between SpotTune and standard fine-tuning narrowed, indicating reduced need for extensive network updates.
• The model achieved a ROC AUC of ~0.9 on FFDM patches, demonstrating strong baseline performance on the source domain.
• The results show that domain shift reduction via preprocessing (HM) is more effective than complex fine-tuning strategies when target data is limited.
• The study demonstrates that effective adaptation to new imaging modalities is possible even with minimal labeled data in the target domain.