# Deep Learning Enhances Weightbearing CT Detection of Lisfranc Instability: A FIXUS-AI Ankle Insight 3D Algorithm

## TL;DR

This study shows that deep learning models can accurately detect Lisfranc instability in weightbearing CT scans, potentially improving diagnosis of subtle foot injuries.

## Contribution

A novel deep learning algorithm (FIXUS-AI Ankle Insight 3D) was developed and tested for detecting Lisfranc instability with high accuracy.

## Key findings

- A differential CNN-LSTM model achieved an F1-score of 0.99 in detecting Lisfranc instability.
- DL models demonstrated significantly higher diagnostic accuracy compared to traditional methods.
- The study found no significant baseline differences between case and control groups.

## Abstract

Background

Recent deep learning (DL) techniques have demonstrated multiple breakthroughs in improving the detection of musculoskeletal pathologies through clinical imaging. Weightbearing CT (WBCT) has been shown to improve diagnostic accuracy in Lisfranc instability, particularly when it is subtle. The aim of the present study was to investigate the impact of applying DL algorithms on WBCT images for the diagnosis of isolated Lisfranc instability.

Methods

The WBCT scans of 280 patients were evaluated (140 cases who had isolated Lisfranc instability, 140 controls without any foot injuries). The entire data set in each group was divided into the training set, validation set, and test set with an 80:10:10 split ratio, in a random manner. Three DL models were developed: (1) a 3D convolutional neural network (3D-CNN); (2) a CNN with long short-term memory (LSTM); and (3) a differential CNN-LSTM. After training, the models’ performance was assessed by means of sensitivity, specificity, accuracy, F1-score, and the area under the receiver operating characteristic (ROC) curve.

Results

The case group included 41% males, and the control group 43%. Mean age and BMI were 35.7 and 26.6, respectively, in the case group, and 32.6 and 27.1 in controls. No significant baseline differences were found. Model 1 had an F1-score of 0.72, while Models 2 and 3 demonstrated substantially higher F1-scores of 0.92 and 0.99, respectively.

Conclusion

This study developed a DL model for 3D WBCT-based Lisfranc injury detection with excellent accuracy. The findings suggest that DL integration has the potential to improve diagnostic accuracy. Further research should focus on larger datasets and external validation.