Sovereign 8s.is mirror of standardgalactic/MSR-AN
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-04-30 13:58:07 +05:00
Code Create model_1_MSA_SN.mat 2026-04-30 13:26:48 +05:00
Images Add files via upload 2025-01-06 17:09:52 +04:00
.DS_Store Create .DS_Store 2026-04-30 13:21:59 +05:00
README.md Update README.md 2026-04-30 13:58:07 +05:00

Synergistic Fusion of a Multilevel Visual Transformer in CNN for Variable-Length Volumetric Radiographic Data Analysis and Content-based Retrieval

Authors: Muhammad Owais, Muhammad Zubair, Taimur Hassan, Divya Velayudhan, Naoufel Werghi, and Irfan Hussain


Overview View the Demo

Dense Crop Detection
Figure 1: Overview of the Synergistic Fusion Framework

This repository showcases the implementation of a novel deep learning framework designed for variable-length volumetric radiographic data analysis in biomedicine. Key innovations include:

  • A multilevel spatio-recurrent attention network (MSR-AN).
  • Efficient integration of a lightweight vision transformer with a convolutional neural network (CNN).
  • Capability to analyze variable-length 3D inputs using single annotations.

View the Demo | View the Code


Dataset

Dataset Samples
Figure 2: Samples from the Volumetric Radiographic Datasets

Explore the datasets used in our evaluation:

  1. Dataset A - Lung CT Scans
  2. Dataset B - Public Radiographic Repository
  3. Dataset C - Composite Repository

Results

Performance Comparison
Figure 3: Performance Comparison of Proposed Model with Baselines

Metric MSR-AN (Proposed) Baseline Model Improvement
Accuracy (%) 98.54 93.46 +5.08%
F1 Score (%) 98.51 93.33 +5.18%
Avg. Precision (%) 98.77 94.12 +4.65%
Avg. Recall (%) 98.25 92.59 +5.66%

Methodology

The proposed framework integrates:

  1. Multilevel Spatio-Attention Sub-Network (MSA-SN):

    • Extracts spatial features using a synergistic fusion of a lightweight vision transformer and CNN.
    • Employs attention mechanisms for multilevel feature aggregation.
  2. Recurrent Feature Alignment Sub-Network (RFA-SN):

    • Processes variable-length sequences using self-attention and LSTM.
    • Captures 3D structural dependencies for volumetric analysis.
  3. Two-Step Training Strategy:

    • Sequentially trains MSA-SN and RFA-SN using cross-entropy loss.
    • Freezes intermediate layers to stabilize learning.

For details, refer to Methodology Section.


Applications

  • Computer-aided diagnosis for pulmonary diseases.
  • Analysis of volumetric radiographic data (e.g., CT, MRI).
  • Multimodal data integration for enhanced medical decision-making.

Citation

If you use this code or dataset, please cite:

@article{owais2024synergistic,
  title={Synergistic Fusion of a Multilevel Visual Transformer in CNN for Variable-Length Volumetric Radiographic Data Analysis and Content-based Retrieval},
  journal={Preprint},
  year={2024},
}