Guidance on ONNX workflow for echocardiography (PSAX AV/PV recognition)

General Discussion
,
1

Hello Hailo Community,

I am working with the following setup and would appreciate guidance on the correct workflow.

Current hardware / environment

  • Raspberry Pi 5

  • Raspberry Pi AI Kit (M.2 HAT+ with Hailo AI Module, Hailo-8 family)

  • Camera pipeline is already working

  • Object recognition with camera input is functional on the Raspberry Pi

  • Development host: Ubuntu 24.04 x86_64 with Docker-based Hailo AI Software Suite

What is working

  • The Hailo device is recognized

  • Camera-based object detection examples run successfully

  • ONNX → HEF compilation flow has been verified with simple test models

What I am trying to do
I am an echocardiographer and would like to test whether Hailo can distinguish anatomical structures in echocardiography.

Specifically:

  • PSAX (parasternal short-axis) echocardiography images

  • Classification or recognition of

    • AV (aortic valve)

    • PV (pulmonary valve)

The initial goal is not clinical-grade performance, but to verify whether this anatomical distinction is feasible on Hailo.

My understanding so far

  • To run a custom model on Hailo, an ONNX model is required as input to the Hailo compiler

  • The ONNX model must be created by training a neural network (e.g., classification or detection) on labeled images

  • Hailo itself does not provide medical or echocardiography-specific ONNX models

Questions

  1. Is this understanding correct that a custom-trained ONNX model is required for this use case?

  2. For a first attempt, would frame-level classification (PSAX_AV vs PSAX_PV) be the recommended approach over detection/segmentation?

  3. Are there any reference ONNX models, operator constraints, or best practices recommended for Hailo when training and exporting ONNX models for custom image classification tasks?

  4. Are there known limitations or considerations when using grayscale or ultrasound-like images with Hailo?

Any guidance, documentation pointers, or practical advice from similar use cases would be greatly appreciated.

Thank you in advance.

1 Like
2

Hi @user517 ,

Thanks for sharing this extremely interesting use case, and we are happy you’re exploring Hailo for that.

My understanding so far: You are correct.

Question 1: It my might be the case.

Question 2: Frame level will do the boolean object detection (detected or not and if so bounding box). Hailo advantage is the exceptional FPS rate, specifically allowing in this case a time series kind of analysis - to track (we have an object tracking capability) the object detection in consecutive frames with a minimal time gap between them, and then a user algorithm can classify the phenomena etc. logic.

Question 3: The supported operations by the parser are in the DFC guide in the developer zone (registration is free): Log In • Hailo

Question 4: Generally speaking there suppose to be no special issues with gray-scale.

Generally: I would recommend to install our official example applications repo and try applying time series like analysis via the callback functions:

Example callback: