How to apply MaskRCNN for Medical Images: Explanations and Code

In this video we learn about MaskRCNN and its usage in Medical Imaging.

Some existing works on the topic are:

Most existing solutions are in Tensorflow 1 which poses difficulty for fine tuning.

[rashmininayanathara]

Thank you, mam. I had a big problem with finding compatible versions of Keras and TensorFlow.
but after watching this video I solved that problem within a few minutes...Thank you again ....

[KrithikaAnbuDeviMPHD]

hai mam, you clearly explained everything, thank you

[saqibqamar]

Hi Sohini,
Excellent explanation to implement MaskRCNN with custom dataset. Pls, let us know how we can calculate the accuracy of generated mask instead of the mAP of object detection? What would be the best metric to evaluate the instance segmentation, not detection?

[RS-vwny]

Hi, thank you for the video, everything was clearly explained. Just have a brief question:

I have some images from centrioles (they are some components of cells important for cell division) and my objective is to identify every centriole. However, for training the model, it would be useful to give an input of more than one image (instead of one image per input, which is the common thing to do, just like you did in your example). This happens because for identifying centrioles we use two images, the first one identifies with a green marker one protein of the centrioles (CEP135), and the other picture identifies another protein with a red marker (Centrin).

I can merge both images into one, but the identification becomes much more difficult to do, especially when centrioles overlap (which happens regularly in the images). It's easier for a human to identify how many centrioles there are when he has the two images separate. My question is: do you know if it's possible to feed more than one image for training sample? Do I have to change the MaskRCNN architecture? Is there another model prepared for this type of input?

Thank you and best regards.

[saqibqamar]

Hi Sohini,
I have trained MaskRNN model with my custom dataset. Could you tell me about how to measure large image where few objects span to another patch. In that case, objects metrics are not accurate.

[vishrutjain]

hi, thank you so much for this video.
I wanted to ask if there is a way to use Mask RCNN pretrained model with .nii files (MRI scans of heart). The dataset itself is already annotated but both the images and the annotations are in .nii format.

[saqibqamar]

Hi Sohini,
Thank you for your video. I am interested to do instance segmentation of wood data that associates with 3 classes including background. In the dataset, few object regions in the classes are overlapped. My questions are

1- Can I use MaskRCNN in my problem?

2- How can I measure the overlapped region of classes? Or the quantitative analysis will give the actual output as Compre to qualitative analysis (visual)..pls reply

[emanwaqar]

Hi Sohini,

I am trying to segment lung x-rays using maskRCNN. I am new to this and having several issues.

Please let me know how you created the masks for training? I am trying to find a way to do that but can't find it.

It will be of great help. Thanks.

[wakandavernon]

is there any video on implementing maskrcnn on custom dataset?

[vivekanandchimman]

Great!
Hi Sohaniji
How can I do it on video dataset for? (Eg:vehicle classification)
I mean how I can detect different types of object (eg: vehicles) in a single video.

[AmadeaPermanaS]

thank you for your explanation it's very helpfull, but I want to ask about the evaluation part. Did mAP is the only evaluation that can evaluate when implementing mask rcnn?

[hadjersaidi]

Hi Sohini
Thank you very much for this meaninggful explanation . I have few questions please
1. How did you create the Masks in the treaning file ?
2. Is there a way to use this code for .jason annotations?

[tiomixt]

Hey thanks for this video. Unfortunately, when I try to run the %tensorflow_version cell in Colab it’s written « tensorflow 1 is unsupported in Colab ». Do you know how to fix this ?

[umarsadique]

how I can apply mask R-CNN to the SVS files of size 0.5GB to 3.5GB. I have data from the Microscope. I am interested to collect two different types of nuclei. How I can collect this data using Mask-RCNN.

2nd question: is instance segmentation is enough for this task or I have also to do first semantic segmentation and then instance segmentation.

3rd question is: If I apply instance segmentation. I worried it may be given me more than two classes but I am interested only in two classes.



thank you for your precious time.

[irfanullah]

Hi Mem Thank you please One video on Faster RCNN in Brain tumor Data set

[mardeenosman]

Hi Sohini
Thank you thank you thank you for amazing videos. I have few questions please
1. Can I use mask RCNN for DICOM images?
2. How would the data look like after preparing? You know its already at least 600 images and there would be masks for every object in all images?? I would love to see a sample please.
Thanks alot for your time

[lijinp]

I'm getting the following error"could not broadcast input array from shape (896, 896, 3) into shape (896, 896)".In my dataset, images are RGB and masks also RGB.how to change the code in the load mask section

[vivekanandchimman]

Hi Sohaniji.
I am getting error ' importErrro: can not import name 'saving' from keras.engine' while loading weights. please help to overcome this.

[Ccomputervision]

Thank for u video. I have a question how to create data mask in video 14:02

[mohammedfadel]

Hi
Can I apply mask rcnn with transfer learning for images classification?