TL;DR
This paper introduces a deep learning model (ResNet) that significantly outperforms experienced ultrasound physicians in differentiating between benign fibroadenomas (FAs) and potentially malignant phyllodes tumors (PTs) using ultrasound images. This advancement holds promise for improving diagnostic accuracy and guiding more precise surgical treatments for breast lesions.
What Problem Does This Paper Address?
Accurately distinguishing between breast fibroadenomas (FAs) and phyllodes tumors (PTs) is crucial for determining the appropriate treatment. While both are fibroepithelial tumors, FAs are typically benign and can be managed non-surgically, whereas PTs can be locally invasive and require wider surgical margins to prevent recurrence. Misdiagnosis can lead to unnecessary surgeries or inadequate treatment. Current diagnostic methods, including ultrasound, often struggle with reliable differentiation due to the visual similarities between these lesions. While AI has shown potential, previous studies have been limited by small sample sizes, weakening the confidence in their findings.
Methods
The researchers employed three popular deep learning architectures: ResNet, VGG, and GoogLeNet. These models were trained on ultrasound images to classify lesions as either FA or PT. The process involved:
- Image Preprocessing: Raw DICOM images were converted to JPG, and regions of interest (ROIs) containing the tumors were manually cropped. Image enhancement techniques were applied to improve contrast and detail.
- Dataset Division: The dataset was split into training, validation, and testing sets. A five-fold cross-validation strategy was used to ensure robust model evaluation.
- Model Training: The models were trained using a cross-entropy loss function. Specific parameters like image size (224x224 pixels) and batch size (16) were optimized.
- Performance Evaluation: The diagnostic performance of the deep learning models and human experts was assessed using metrics such as Area Under the Curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
Datasets
The study utilized a retrospective dataset of 656 breast lesions from 656 female patients, comprising 372 fibroadenomas (FAs) and 284 phyllodes tumors (PTs). This dataset included a total of 1945 ultrasound images. The PTs were further categorized into benign (134), borderline (120), and malignant (30) subtypes. This dataset is notably larger than those used in previous deep learning studies for this specific differentiation task.
Key Insights and Ideas
- Deep Learning's Superiority: The study strongly suggests that deep learning models, particularly ResNet, can achieve significantly higher diagnostic accuracy than experienced human physicians in differentiating PTs from FAs using ultrasound. This highlights AI's potential to augment, rather than replace, clinical expertise.
- ResNet's Architectural Advantage: The superior performance of ResNet is attributed to its deeper and wider network structure, its use of residual blocks to mitigate degradation issues in deeper networks, and its implementation of batch normalization to address gradient problems.
- Clinical Relevance of Demographics: Incorporating patient age and tumor size into the deep learning model (ResNet) led to a slight improvement in diagnostic performance, underscoring the importance of these clinical factors in differentiation.
- Physician Performance Variability: While experienced, the physicians' performance showed variability, particularly with smaller lesions, where sensitivity dropped significantly. This suggests that subtle visual cues, which AI can process more objectively, are critical.
Key Results
- ResNet Performance: In the testing dataset, the ResNet model achieved an AUC of 0.91, accuracy of 95.3%, sensitivity of 96.2%, and specificity of 94.7%.
- Physician Performance: The two experienced ultrasound physicians achieved an average AUC of 0.69, with an average accuracy of 70.7%, sensitivity of 54.4%, and specificity of 53.2%.
- Comparison: The ResNet model's AUC (0.91) was substantially higher than the physicians' average AUC (0.69). Similarly, ResNet's accuracy (95.3%) far surpassed the physicians' average accuracy (70.7%).
- Impact of Clinical Data: When age and tumor size were included in the ResNet model, the AUC increased to 0.94, and accuracy to 96.4%.
Limitations and Future Work
- Single-Center Retrospective Study: The study was conducted at a single center retrospectively, which may introduce selection bias. External validation with larger, multi-center cohorts is necessary to confirm the model's generalizability and clinical utility.
- Limited Scope: The model is specifically trained for differentiating PTs and FAs. Its applicability to other breast lesions (e.g., invasive cancers, inflammation) is not addressed and requires further investigation.
- Algorithm Comparison: While three common deep learning models were used, the study acknowledges that other advanced algorithms (e.g., deep distance learning) were not explored.
- Direct Comparison of Metrics: The authors note that directly comparing AUC values between deep learning models and physicians can be challenging due to the different underlying algorithms used for their respective assessments.
Future work should focus on external validation, expanding the model's scope to include other breast pathologies, and potentially exploring more advanced deep learning architectures.
Summary updated