Yolo Model Training Discussion

Hello, can anyone help please, I am training a yolov8 model to detect a dental lesion from dental X-ray images. here are some samples of my data:

after training the model I noticed that there is a problem with the sensitivity, so the model miss rate is still high and it drops a lot of detections as you see in these results.

I had a total of 221 images with exactly 1020 target objects in them so I had to do a little bit of image augmentation (only pixel level transforms) like rand histogram equalizing, blurring, and contrast change. From the original image, I got 6 augmented images so the final is 221 + 221*6 = 1547 images with their label text files divided as follows:
train: 60%, val 20%, and test 20%

the hyperparameters for training will be found below:

please if you have any suggestions for better results share them with me.
thank you in advance.

did you label your data for this model?

yup dental pathologist helped me with the data annotation, the sample images above are actually of the labeled images I have.

can I know when you split the dataset, you did random selection?

that part you mentioned you bodied 6 images out of the 221 images !!! and then you added the same dataset to the 221 images with your 221*6 is creating issue.

Have a look at this link, although your data augmentation idea was quite right but you should not have selected 6 images and split the dataset.

Hope it helps!!!


thank you, I will check it and reply to you

okay, I still can not find out where the problem is.
I understand that the more image we get from augmentation the better it will be for training, that is why I created 6 (6 is randomly chosen, no reasons for it) different augmented images from each original image each with different blurring and contrast levels.
The reason why I have chosen pixel level transforms not spatial lever transforms, is that panoramic xrays have standards, so why do I need to train my model on rotated images if I am sure that we will never ask it later to do detection of rotated images.

please note that I am trying to segment the lesions, not just classify them, that is why I think it will need more images to learn.

Hello Zidan,

Based on your post, I think one need to keep improvising the data augmentation part until you want to get better accuracy.

The next issue I noticed is with the splitting of data.

you got final 1547 images and splitting that into 60-20-20 is another issue. Please read the link properly have sent, you will understand. Make the pointers on what difference your model and the link I sent you differs, and then work on it. if you still are getting doubt, let me know.