Hi, everyone,
I need to apply Computer Vision (Convolutional Networks, i think) to find abnormal structures in 3D Medical Images.
Do you have any resource? It can be text or video content, or courses.
I already have experience with Machine and Deep Learning for 2D Images, numerical data and Video data, but i've never worked with 3D medical imaging before. I need to learn how to detect abnormal structures in medical imaging data to, for example, draw a bounding box around the pattern in the image.
I found that Computer-Aided Detection is the technique applied in Medical Imaging that can find abnormal structures, but i still don't know how to do this.
I think i can do this by using tensorflow and some image processing library, like Nibabel, Matplotlib, and others (i have basic knowledge in the field of Computer Vision, and i'm diving even more deeper through the PyImageSearch Computer Vision course)
What i really need is a direction to walk through this learning path, starting from reading MRI data to apply a convolutional neural network to find a pattern in the medical image. The bounding box is a pretty simple task, i think. :)
Remembering, i don't need a tutorial on "how to use tensorflow", for example, don't worry about this, i just want to know how to take, for example, a CNN built on TensorFlow, or Keras, for example and make this network find an abnormal structure on the medical image, and after this, draw the bounding box.
I'm experienced in Python Programming and i need this to start to work in a project. :)
Thanks for the attention!
Suggestions and Reccomendations for the highest fidelity of analysis.
1) KNOW your Data very well and don’t underestimate this activity...know how your Data was collected and as much about the context of when where why and any addl details, detailed context will help you better develop your weights, bias and filters,
2) if you are dealing with imagery of any kind you must understand the camera models for Whatever generated the imagery...example, many biologists lack fidelity upon analysis of phenotypic arrays because they simply scan their microarrays without knowledge of the scanner camera model...if they did it would increase accuracy substantially
3) even ambient temperature can sometimes be impt, if accessible
4) cleanse and tag your data well so you can track it better during the learning process
5) depending on the type of imagery and/or sensor you can generate some depth with multiple stacks of 2D analysis...for Red, Green, Blue features or maybe multiple more for other bands if you need to model higher band sensors
6) the third dimension seems to me would essentially be another feature
7) working through your anomaly detection then will depend upon which classifier you choose and how well you calculate your bias
Hopefully this helps some. Not sure how much of this you already know.
Dear Arthur,
considering your question, I would suppose you're a novice in CNNs. Then there are two different things you should do:
3D image segmentation and classification: We have performed two reviews that will interest you:
Article 3D Segmentation Algorithms for Computerized Tomographic Imag...
Technical Report Literature review for 3D object classification/recognition
Introductory material:
Hi I suggest you to take look on probabilistic models like RBM, DBN of any kind of Autoencoders.
Best
- Badges
- Science topic