TL;DR
This paper introduces a deep learning model (DLM) that accurately diagnoses breast cancer from ultrasound images, outperforming traditional BI-RADS categorization. Crucially, the DLM can also predict molecular subtypes, offering valuable information for treatment planning and potentially reducing unnecessary biopsies.
What Problem Does This Paper Address?
Breast cancer diagnosis heavily relies on ultrasound, particularly in regions like China where breast density is higher. The current standard, the Breast Imaging Reporting and Data System (BI-RADS), is effective but often leads to unnecessary biopsies due to its broad categorization, especially for BI-RADS 4a lesions. Furthermore, radiologists cannot predict crucial molecular subtypes (like hormone receptor and HER2 status) from ultrasound alone, information vital for guiding neoadjuvant therapy and personalized treatment. This gap highlights the need for a more precise and informative diagnostic tool.
Methods
The researchers developed a deep learning model (DLM) using a transfer learning approach. They started with a pre-trained Xception convolutional neural network (CNN) model, initially trained on the massive ImageNet dataset (CNN-1). This model was then fine-tuned for breast cancer diagnosis (CNN-2) and subsequently for molecular subtype classification (CNN-3). The CNN-2 model was trained to distinguish between benign and malignant tumors, while CNN-3 focused on classifying the three main molecular subtypes: HR (+), HER2 (+), and triple-negative. Image preprocessing and data augmentation techniques were employed to enhance image quality and data diversity.
Datasets
The study utilized ultrasound images from two hospitals:
- Harbin Medical University Cancer Hospital: Provided the primary dataset for training and testing.
- The First Affiliated Hospital of Harbin Medical University: Provided an independent external test set for validation.
The datasets were carefully curated, excluding images with incomplete pathological data. The final training set included 2,822 images, the test set had 707 images, and the external test set comprised 210 images. For molecular subtype prediction, specific subsets of these images were used.
Key Insights and Ideas
- DLM Outperforms Radiologists: The deep learning model demonstrated superior diagnostic accuracy compared to radiologists using the BI-RADS system.
- Reducing Unnecessary Biopsies: The DLM showed a significant potential to reduce unnecessary biopsies, particularly for BI-RADS 4a lesions, by more accurately identifying benign cases.
- Predicting Molecular Subtypes: This is a novel application of DLMs in breast ultrasound, enabling the prediction of crucial molecular subtypes directly from imaging data, which is not possible with current radiological methods.
- Transfer Learning Efficacy: The use of transfer learning with a pre-trained Xception network proved effective in achieving high performance with a relatively smaller, specialized dataset.
Key Results
- Diagnostic Accuracy:
- Test Set: DLM achieved an accuracy of 89.7% compared to 85.6% for BI-RADS categorization.
- External Test Set: DLM achieved an AUC of 0.90, while BI-RADS had an AUC of 0.96 on the test set.
- Reducing Biopsies: For BI-RADS 4a patients, the DLM achieved 92.86% diagnostic accuracy, theoretically reducing unnecessary biopsies by 67.86%.
- Molecular Subtype Prediction: The DLM showed promising results in predicting molecular subtypes, with AUCs of 0.864 for triple-negative, 0.811 for HER2 (+), and 0.837 for HR (+) subtypes on the test set.
Limitations and Future Work
The authors acknowledge several limitations:
- Single-Institution Training Data: The primary training set originated from a single hospital, potentially limiting generalizability.
- Small Sample Size: The sample sizes for both training and testing sets, especially the external test set, were relatively small.
- Retrospective Nature: The study was retrospective, and all patients underwent surgical treatment. This might not fully represent cases where observation is chosen over surgery for potentially malignant lesions.
Future work should focus on expanding the dataset by collaborating with multiple institutions to validate the model's performance and clinical utility in a larger, more diverse cohort. Further refinement of the DLM could also enhance its decision-making capabilities for clinical applications.
Summary updated