Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BowelRCNN: Region-based Convolutional Neural Network System for Bowel Sound Auscultation

Igor Matynia, Robert Nowak

TL;DR

The paper tackles automatic detection of bowel sounds from audio to aid gastrointestinal diagnostics. It introduces BowelRCNN, a two-stage region-based CNN that operates on MEL spectrograms to first locate candidate windows and then refine their time intervals, trained and evaluated on a Kaggle dataset of 19 patients. Key contributions include the two-stage detector design, robust training with augmentation and carefully tuned prediction parameters, and a thorough comparison against CRNN baselines, achieving mean IoU around 0.55 and F1 around 0.71 while maintaining high accuracy. The work demonstrates the feasibility of CNN-based bowel sound auscultation with practical performance and suggests directions for real-time deployment and broader data, potentially impacting noninvasive GI diagnostics, especially for patients unable to communicate symptoms.

Abstract

Sound events representing intestinal activity detection is a diagnostic tool with potential to identify gastrointestinal conditions. This article introduces BowelRCNN, a novel bowel sound detection system that uses audio recording, spectrogram analysys and region-based convolutional neural network (RCNN) architecture. The system was trained and validated on a real recording dataset gathered from 19 patients, comprising 60 minutes of prepared and annotated audio data. BowelRCNN achieved a classification accuracy of 96% and an F1 score of 71%. This research highlights the feasibility of using CNN architectures for bowel sound auscultation, achieving results comparable to those of recurrent-convolutional methods.

BowelRCNN: Region-based Convolutional Neural Network System for Bowel Sound Auscultation

TL;DR

The paper tackles automatic detection of bowel sounds from audio to aid gastrointestinal diagnostics. It introduces BowelRCNN, a two-stage region-based CNN that operates on MEL spectrograms to first locate candidate windows and then refine their time intervals, trained and evaluated on a Kaggle dataset of 19 patients. Key contributions include the two-stage detector design, robust training with augmentation and carefully tuned prediction parameters, and a thorough comparison against CRNN baselines, achieving mean IoU around 0.55 and F1 around 0.71 while maintaining high accuracy. The work demonstrates the feasibility of CNN-based bowel sound auscultation with practical performance and suggests directions for real-time deployment and broader data, potentially impacting noninvasive GI diagnostics, especially for patients unable to communicate symptoms.

Abstract

Sound events representing intestinal activity detection is a diagnostic tool with potential to identify gastrointestinal conditions. This article introduces BowelRCNN, a novel bowel sound detection system that uses audio recording, spectrogram analysys and region-based convolutional neural network (RCNN) architecture. The system was trained and validated on a real recording dataset gathered from 19 patients, comprising 60 minutes of prepared and annotated audio data. BowelRCNN achieved a classification accuracy of 96% and an F1 score of 71%. This research highlights the feasibility of using CNN architectures for bowel sound auscultation, achieving results comparable to those of recurrent-convolutional methods.

Paper Structure

This paper contains 14 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Bowel sounds analysis
  3. Methods: BowelRCNN - the bowel sound detection system
  4. Signal preprocessing
  5. Bowel sound detector
  6. Training the convolutional neural networks
  7. Predictions aggregation
  8. Results
  9. Model training parameters optimization
  10. Tweaked network architecture
  11. Increased spectrogram size experimentation
  12. Prediction parameters optimization
  13. Overall model comparison results
  14. Summary

Figures (3)

  • Figure 1: The figure shows the general overview of initial bowel sound recording processing
  • Figure 2: Bowel sound detector. This diagram excludes the initial data processing and predictions aggregation
  • Figure 3: The heatmap shows the values of selected metrics as well as selected detection threshold values on the horizontal axis. The brighter the color of the cell, the higher the value. The values have been averaged over 5 models trained with different random number generator seeds.