Introduction
Peritoneal metastasis (PM) represents a significant challenge in the management of gastric cancer. The current diagnostic standard involves the cytological examination of ascites to identify free-floating malignant cells. However, this manual screening process is labor-intensive and prone to human error due to the "needle in a haystack" nature of finding sparse cancer cells among thousands of benign ones. This study introduces a two-stage deep learning (DL) framework designed to automate and enhance the precision of ascites cytopathology interpretation.
System Architecture
The researchers developed an end-to-end system comprising two specialized modules: DetectionNet and ClassificationNet. This decoupled approach allows for modularity and targeted optimization of each task.
DetectionNet
The first stage, DetectionNet, utilizes the Faster R-CNN architecture with a ResNet18 backbone. Its primary function is to perform cell localization, identifying and bounding every individual cell within a high-resolution cytopathology image.
ClassificationNet
Following detection, the ClassificationNet (based on ResNet50) analyzes the localized cell patches to differentiate between benign and malignant morphologies. To ensure high sensitivity to subtle features like nuclear-cytoplasmic ratios, the system employs four-fold magnification via bicubic interpolation before classification.
Training and Clinical Application
The system leverages transfer learning, initially training on the ImageNet dataset before fine-tuning on a specialized medical dataset of 139 patients. This dataset included both Hematoxylin-Eosin (HE) and Papanicolaou (PAP) stained images, totaling over 25,000 annotated cells.
The proposed workflow envisions the DL system as a high-speed "pre-screener." Capable of processing approximately 100 cells in 3.7 seconds, the system flags suspicious cells for expert confirmation, thereby reducing the diagnostic burden on pathologists without sacrificing clinical oversight.
Key Results
The validation of the system yielded impressive performance metrics:
- Detection Precision: DetectionNet achieved a mean Average Precision (mAP) of 0.8316.
- Classification Accuracy: ClassificationNet reached a precision of 96.80% with an AUC of 0.8851.
- Clinical Reliability: The system maintained a low False Negative Rate (FNR) of 4.73%, crucial for ensuring that no malignant cases are missed during screening. At the image level, the system achieved a precision of 97.5%.
Discussion and Future Directions
While the results are promising, the study identifies several areas for future development. The current model was trained on data from a single center, necessitating multi-center validation to account for staining and equipment variability. Additionally, the system currently excludes cell clusters—a common occurrence in malignant samples—which will be addressed in future iterations to improve robustness in real-world clinical settings.
The decoupling of detection and classification not only improves performance but also offers scalability; the ClassificationNet could potentially be adapted for other cancer types with minimal adjustments to the detection layer.
Summary updated