MM-SFENet: multi-scale multi-task localization and classification of bladder cancer in MRI with spatial feature encoder network

Yu Ren, Guoli Wang, Pingping Wang, Kunmeng Liu, Quanjin Liu, Hongfu Sun, Xiang Li, Benzheng Wei

Abstract

Bladder cancer (BC) staging, specifically distinguishing between Non-Muscle-Invasive (NMIBC) and Muscle-Invasive (MIBC), is critical for determining treatment paths. While MRI is a non-invasive alternative to cystoscopy, automated staging is challenging due to low contrast and complex spatial relationships. This paper presents the Multi-scale Multi-task Spatial Feature Encoder Network (MM-SFENet), an end-to-end deep learning model designed to locate and classify BC subtypes. By employing a specialized Spatial Feature Encoder (SFE) and IoU-based loss functions, MM-SFENet achieves a mean Average Precision (mAP) of 93.34%, demonstrating significant potential for non-invasive clinical diagnosis.

Introduction

Bladder cancer remains a major global health challenge. The distinction between NMIBC and MIBC is the primary factor in deciding between local resection and radical surgery. Currently, optical cystoscopy is the gold standard but is invasive and painful. MRI offers high soft-tissue contrast but requires expert interpretation to identify subtle spatial cues indicating muscle invasion. Standard neural networks often struggle with the "needle in a haystack" problem in MRI due to low contrast between tumors and the bladder wall, as well as the high variability in tumor morphology. MM-SFENet addresses these challenges by moving beyond simple segmentation to an integrated localization and classification approach.

Methods

The MM-SFENet architecture is built on three core components designed to bridge the gap between low-level spatial features and high-level semantic information.

  1. Backbone Architecture: The model utilizes a deep CNN (e.g., ResNet50) with residual connections to extract robust semantic features while avoiding the vanishing gradient problem.
  2. Spatial Feature Encoder (SFE): The SFE is a critical innovation that fuses high-resolution spatial features from early layers with low-resolution semantic features from deeper layers. This multi-scale fusion allows the network to better perceive the spatial relationship between the tumor and the bladder wall, which is essential for staging.
  3. IoU-based Multi-task Learning: Unlike standard detectors that use Smooth-L1 loss for bounding boxes, MM-SFENet employs Intersection over Union (IoU) loss. This treats the bounding box as a single unit, ensuring scale consistency which is vital for accurate staging in medical imaging where small errors can lead to misdiagnosis.
Overview of the MM-SFENet architecture, showing the pipeline from input MRI to final classification and localization.
Detailed view of the Backbone and Encoder (SFE) components, illustrating the feature fusion mechanism.

Additionally, the authors optimized anchor box sizes using k-means++ clustering on their specific dataset to better match the morphology of bladder tumors compared to standard anchors designed for natural images.

Results

The model was evaluated on a dataset of 1,287 MRI images from 98 patients, annotated by expert radiologists. MM-SFENet achieved a mAP of 93.34% and a localization accuracy (IoU) of 83.16% on the test set.

Key findings from the study include:

  • Ablation Success: The inclusion of the SFE component alone contributed a 20% increase in mAP over the baseline ResNet model, proving the necessity of multi-scale feature fusion.
  • Superior Performance: MM-SFENet outperformed several state-of-the-art models, including Cascade R-CNN, YOLOv4, and SSD.
  • Interpretability: Grad-CAM++ visualizations indicated that the model focuses on the continuity of the bladder wall, aligning with the diagnostic logic used by human radiologists.

Conclusion

MM-SFENet provides a powerful, non-invasive method for bladder cancer staging. By effectively encoding spatial relationships and optimizing for medical-specific morphology, it achieves high precision in distinguishing between NMIBC and MIBC. While currently a supplementary tool, its performance marks a significant step toward automated MRI-based diagnosis. Future work will focus on more granular clinical staging (e.g., T1-T3) and validation on larger, multi-center datasets to ensure broader clinical applicability.

Summary updated

Abstract

<i>Objective</i>. Bladder cancer is a common malignant urinary carcinoma, with muscle-invasive and non-muscle-invasive as its two major subtypes. This paper aims to achieve automated bladder cancer invasiveness localization and classification based on MRI.<i>Approach</i>. Different from previous efforts that segment bladder wall and tumor, we propose a novel end-to-end multi-scale multi-task spatial feature encoder network (MM-SFENet) for locating and classifying bladder cancer, according to the classification criteria of the spatial relationship between the tumor and bladder wall. First, we built a backbone with residual blocks to distinguish bladder wall and tumor; then, a spatial feature encoder is designed to encode the multi-level features of the backbone to learn the criteria.<i>Main Results</i>. We substitute Smooth-L1 Loss with IoU Loss for multi-task learning, to improve the accuracy of the classification task. By learning two datasets collected from bladder cancer patients at the hospital, the mAP, IoU, Acc, Sen and Spec are used as the evaluation metrics. The experimental result could reach 93.34%, 83.16%, 85.65%, 81.51%, 89.23% on test set1 and 80.21%, 75.43%, 79.52%, 71.87%, 77.86% on test set2.<i>Significance</i>. The experimental result demonstrates the effectiveness of the proposed MM-SFENet on the localization and classification of bladder cancer. It may provide an effective supplementary diagnosis method for bladder cancer staging.
BibTeX
@inproceedings{ren2023mm,
  author = {Yu Ren and Guoli Wang and Pingping Wang and Kunmeng Liu and Quanjin Liu and Hongfu Sun and Xiang Li and Benzheng Wei},
  title = {{MM-SFENet: multi-scale multi-task localization and classification of bladder cancer in MRI with spatial feature encoder network}},
  booktitle = {Physics},
  year = {2023},
  doi = {10.1088/1361-6560/ad1548},
}
Figure