# MedYOLO: A Medical Image Object Detection Framework

**Authors:** Joseph Sobek, Jose R. Medina Inojosa, Betsy J. Medina Inojosa, S. M. Rassoulinejad-Mousavi, Gian Marco Conte, Francisco Lopez-Jimenez, Bradley J. Erickson  
**arXiv:** [2312.07729](https://arxiv.org/abs/2312.07729)  
**Date:** 2023-12-12  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2312.07729v1/

## Key points

- Artificial intelligence is crucial in medical imaging for identifying organs, lesions, and structures.
- Convolutional neural networks (CNNs) are commonly used for voxel-accurate segmentations.
- Object detection models offer an alternative to reduce annotation effort, especially when voxel-level precision is not necessary.
- MedYOLO is a 3-D object detection framework designed for medical imaging applications using the one-shot detection method from the YOLO family of models.
- MedYOLO showed high performance in detecting medium and large-sized structures like the heart, liver, and pancreas without hyperparameter tuning but faced challenges with very small or rare structures.
- One-shot anchor-based approaches demonstrate effectiveness in accurate 3-D medical object detection.
- Future frameworks could potentially improve by adopting a 2.5-D paradigm using YOLO-like approaches to enhance performance in detecting complex structures while optimizing efficiency in medical imaging applications.

## Summary

In the field of medical imaging, artificial intelligence plays a crucial role in identifying organs, lesions, and other structures. One method commonly used for voxel-accurate segmentations is convolutional neural networks (CNNs). However, generating labels for training these models is time-consuming and requires expertise to ensure quality. To address this issue, object detection models offer an alternative that can reduce annotation effort. This is especially useful for tasks where voxel-level precision is not necessary. MedYOLO is a 3-D object detection framework specifically designed for medical imaging applications. It utilizes the one-shot detection method from the YOLO family of models. The framework was tested on various datasets including BRaTS, LIDC, an abdominal organ Computed Tomography (CT) dataset, and an ECG-gated heart CT dataset. The results showed high performance in detecting medium and large-sized structures such as the heart, liver, and pancreas without hyperparameter tuning. However, challenges were encountered when detecting very small or rarely present structures. Despite its limitations in handling small or uncommon structures, MedYOLO demonstrates the effectiveness of one-shot anchor-based approaches in achieving accurate 3-D medical object detection. In the future, frameworks could potentially improve by adopting a 2.5-D paradigm using YOLO-like approaches to maintain native resolution without compromising batch size or introducing distortion from reshaping. This shift could enhance performance in detecting complex structures while optimizing efficiency in medical imaging applications.