I’m currently training a ML model that can estimate sex based on dimensions of proximal femur from radiographs. I’ve taken x-ray images from ALL of the samples in the osteological collection in Chiang Mai, left side only, which came to a total of 354 samples. I also took x-ray photos of the right femur and posterior-anterior view of the same samples (randomized, and only selective few n=94 in total) to test the difference, dimension wise. I have exhausted all the samples for training the model and validating (5-fold), which results in great accuracy of sexing. So, I am wondering whether it is appropriate to test the models with right-femur and posterior-anterior view radiographs, which will then be flipped to resemble left femur x-ray images, given the limitations of our skeletal collection?
I would not recommend adjusting images of right femur to make into left femur as discriminating features may be then mapped amongst both classes. However, I would recommend trying image augmentation on left femur to balance out the dataset a bit more. There are also sampling methods over/under sampling methods of minority/major classes that aid imbalanced data.
It depends on whether the results of image identification in the software system are invariant with respect to mirror reflection, rotation, scaling and image transfer
Eugene Veniaminovich Lutsenko Well, ideally, the ML algorithm proposed here will be developed as a web-based application, and in case of absent left femur, right femur can be potentially substituted for sex classification. I hope I am making sense. I’ll be willing to clarify more in direct message.
Patara Rattanachet!
It's a good idea. But it is better to use a coordinate system and a unit of measurement associated with the images themselves, rather than a frame or a border of the image.
Eugene Veniaminovich Lutsenko Agreed. Our automated method relies on Canny edge detection technique to measure the features of the femur itself, rather than a frame.
Patara Rattanachet!
https://perso.limsi.fr/vezien/PAPIERS_ACS/canny1986.pdf
https://habr.com/ru/post/114589/
There is still such an idea.
Train the system simultaneously with recognition or clustering. The learning algorithm is as follows. Cycle through images. We take another image of the training sample belonging to a certain class. We identify (recognize) it in the existing model. If the level of similarity of this image with the generalized image of the class to which it really belongs according to the training sample is higher than a certain threshold set by you in the dialog, then this image is used to continue the formation of this generalized image of the class. If the similarity is below the threshold, then a new class is formed, which is also called, but with the addition of some number or prefix. As a result, we will have new classes for the cases of inverted images and images that have undergone other transformations of mirror reflection, displacement (shift), rotation in plane and space, and scaling.
I have not implemented this algorithm in my Eidos system. But I have implemented a divisional clustering algorithm. This algorithm allows you to divide classes into typical and atypical parts. So inverted images will lead to false negative decisions. After divisional clustering, separate classes will be created for them and they will be identified truly positively.
Есть еще такая идея.
Обучать систему одновременно с распознаванием или кластеризацией. Алгоритм обучения такой. Цикл по изображениям. Берем очередное изображение тренировочной выборки, относящиеся к определенному классу. Идентифицируем (распознаем) его в имеющейся модели. Если уровень сходства данного изображения с обобщенным изображением класса, к которому оно действительно относится по данным тренировочной выборки, выше определенного заданного вами в диалоге порога, то это изображение используется для продолжения формирования данного обобщенного образа класса. Если же сходство ниже порога, то формируется новый класс, который называется также, но с добавлением какого-то номера или префикса. В результате у нас получатся новые классы для случаев перевернутых изображений и изображений, подвергшимся другим преобразованиям зеркального отражения, смещения (сдвига), поворота в плоскости и пространстве и масштабирования.
Я этот алгоритм не реализовал в своей системе Эйдос. Но я реализовал алгоритм дивизионной кластеризации. Этот алгоритм позволяет разделить классы на типичную и нетипичную части. Так перевернутые изображения приведут к ложным отрицательным решениям. После дивизионной кластеризации для них будут созданы отдельные классы и они будут идентифицированы истинно положительно.
Eugene Veniaminovich Lutsenko Thank you so much for taking your time to write an in-depth response! Definitely helpful to my research.
Patara Rattanachet!
Here are some of my works that mention the method of cognitive divisional clustering developed and programmatically implemented by me:
https://www.researchgate.net/publication/360176570
https://www.researchgate.net/publication/359369213
https://www.researchgate.net/publication/360474409
https://www.researchgate.net/publication/351924470
https://www.researchgate.net/publication/353562220
Eugene Veniaminovich Lutsenko thank you! I will give those a read!
- Badges
- Science topic
- Similar topics
- Correction