Establishment and Clinical Application Value of an Automatic Diagnosis Platform for Rectal Cancer T-Staging Based on a Deep Neural Network

0
0
0

Establishment and Clinical Application Value of an Automatic Diagnosis Platform for Rectal Cancer T-Staging Based on a Deep Neural Network

Colorectal cancer remains a global health challenge, with its incidence ranking third among malignant tumors and mortality fourth among cancer-related deaths. Accurate preoperative staging is critical for determining treatment strategies, as mismanagement due to inaccurate staging can lead to irreversible consequences. Magnetic resonance imaging (MRI) has emerged as the gold standard for preoperative rectal cancer staging due to its superior soft tissue resolution compared to endoscopic ultrasound (EUS) and computed tomography (CT). However, variability in radiologists’ expertise, high workload, and human error limit diagnostic consistency. This study addressed these challenges by developing an artificial intelligence (AI)-powered platform using Faster Region-Based Convolutional Neural Networks (Faster R-CNN) to automate T-staging of rectal cancer from MRI data.

Methodology and Platform Design

The study retrospectively analyzed 183 rectal cancer patients treated at the Affiliated Hospital of Qingdao University between July 2016 and July 2017. Inclusion criteria required preoperative high-resolution pelvic MRI, postoperative pathological confirmation of T-stage, and radical surgical resection. Patients with ambiguous diagnoses or non-malignant lesions were excluded. The cohort included 121 males (66.1%) and 62 females (33.9%), with tumor locations distributed as 59 lower rectal, 91 mid-rectal, and 33 upper rectal cancers. MRI scans utilized a GE Signa 3.0T HDX scanner with T2-weighted sequences in sagittal, coronal, and horizontal planes. Key parameters included repetition times of 2000–4000 ms, echo times of 60–120 ms, and slice thicknesses of 5–6 mm. A total of 10,800 MRI images were compiled, with 80% (8,640 images) allocated for training and 20% (2,160 images) for validation.

The Faster R-CNN architecture was selected for its ability to simultaneously propose regions of interest (ROIs) and classify tumor stages. The network comprised four components:

  1. Convolutional Layers: Extracted features from input images resized to 512×557 pixels and normalized to standard distributions.
  2. Region Proposal Network (RPN): Generated candidate tumor regions using anchor boxes scaled to rectal anatomy.
  3. ROI Pooling: Standardized variable-sized regions into fixed-length vectors for classification.
  4. Classification and Regression: Predicted T-stages (T1–T4) based on tumor invasion depth.

Training involved a four-step iterative process:

  1. RPN Training: Initialized with VGG16 weights to propose regions using annotated T-stage labels.
  2. Detection Network Training: Used RPN proposals to train stage classification, initialized separately with VGG16.
  3. RPN Fine-Tuning: Fixed convolutional layers while refining RPN parameters using the detection network.
  4. Detection Network Fine-Tuning: Final tuning of classification layers with fixed convolutional weights.

The loss function combined classification and localization errors, decreasing steadily over 700 epochs, indicating successful convergence.

Performance Evaluation

The platform demonstrated exceptional performance across all MRI planes and T-stages:

  1. Plane Identification: Achieved 100% accuracy in distinguishing sagittal, coronal, and horizontal MRI planes.
  2. Overall Diagnostic Accuracy:
    • Horizontal Plane: Area Under the Curve (AUC) = 0.99
    • Coronal Plane: AUC = 0.98
    • Sagittal Plane: AUC = 0.97
  3. Stage-Specific Performance:
    • Horizontal Plane: AUCs for T1, T2, T3, and T4 stages were all 1.00.
    • Coronal Plane: AUCs ranged from 0.96 (T1) to 0.97 (T2–T4).
    • Sagittal Plane: AUCs were 0.95 (T1), 0.99 (T2), 0.96 (T3), and 1.00 (T4).

These results surpassed the 62% diagnostic accuracy of radiologists at the participating hospital and outperformed manual MRI staging accuracy (86%). The platform’s ability to delineate tumor boundaries and classify stages is illustrated in comparative figures showing expert annotations versus AI predictions.

Clinical Implications and Limitations

The study highlights three critical advancements:

  1. Automated Multi-Plane Integration: By analyzing three orthogonal planes simultaneously, the platform compensates for single-plane limitations, improving staging reliability.
  2. Pathology-Driven Learning: Training on postoperative pathological “gold standards” minimized discrepancies between imaging and histopathological findings.
  3. Operational Efficiency: The system processes images faster than manual evaluation, addressing radiologist shortages in high-demand settings.

However, limitations include:

  • Patient Selection Bias: Excluding non-surgical candidates (e.g., metastatic or early-stage endoscopic resection patients) may reduce generalizability.
  • Post-Chemoradiation Restaging: The platform was not tested on post-neoadjuvant therapy cases, where fibrosis complicates MRI interpretation.
  • Class Imbalance: Uneven distribution of T-stages (e.g., fewer T1 cases) could introduce classification bias.

Conclusion and Future Directions

This study successfully established an AI platform for rectal cancer T-staging with diagnostic parity to expert radiologists. By leveraging deep learning and multi-plane MRI analysis, the system provides rapid, consistent staging to guide treatment decisions. Future work will focus on:

  1. Expanding training datasets to include non-surgical and post-chemoradiation cases.
  2. Integrating lymph node (N) and metastasis (M) staging modules.
  3. Validating performance across diverse populations and imaging protocols.

The platform exemplifies AI’s transformative potential in oncology, bridging gaps between imaging resources and clinical demand while reducing diagnostic variability.

doi.org/10.1097/CM9.0000000000001401

Was this helpful?

0 / 0