Is it possible to recognize the injured areas of the hand and choose the appropriate incision on the arm?

Question

Android

Good day, I'm new to machine learning. Perhaps the task is to recognize injuries in a certain area of ​​​​the hand (for example, a certain phalanx of the finger), and draw a suitable incision.
Subtasks come from this:
1. Recognition of a problematic brush and not
2. Recognition of a separate section of a problematic brush
3. Selection of a cut for this problematic section of the network

1. Questions arise, is it possible to solve this problem using machine learning (possibly specific subtasks)?
2. Are there any ready-made tools for this, automatic or semi-automatic?
3. Is it possible to use these models on the Android operating system?
4. If the task can be done, what is worth reading or watching?

Answer 1

1. Yes
2. Only libraries for building systems
3. Yes
4. Everything you can find on machine learning

Might be a task

Maybe not worth it?

Answer 2

Everything is yes, but most likely it will not be possible to find ready-made tools for a specific task, what is walking on the Internet is good for self-study, but for business you still have to tinker with a specific task and its features.
The use of neural networks is not amenable to a clear forecast. It is impossible to say in advance that such a volume of data can be dispensed with for this task .
Learning neural networks is a research task, and if you want to solve a problem in a simple way (without bringing your own knowledge to the algorithms, implementing them in classical deterministic algorithms), then the task is from scratch! which means there should be not just a lot of data,

and even more than it would seem necessary

Например недостаточно сети давать только котиков и собачек, если надо их находить, а возможно придется давать и рыбок, слоников и машинки, чтобы она в принципе могла бы решать задачу, а то найдет ваша сеть в качестве главного признака не форму или пропорции животного, а к примеру угол поворота головы (абстрактный пример - коты чаще будут смотреть в камеру а собаки в бок) или еще хуже, как оператор размечает вокруг животного квадратик, у кошечек левее пикселов чаще больше сделает и сеть тут же на этот принзак сагрится, т.е. дай сети еще и картинки без животного а в выходы добавь и этот результат.

And convey this idea to the customer that neural networks are about data, their correct selection (for example, it is impossible that there is only one black cat in the training sample and all the rest are spotty, the network may not accept this feature and then give such a false refusal) their processing , markup, introduction of distortions and noise (which may appear in production), a typical example of increasing the training sample by rotating each image (for example, all 360 degrees by 2-5 and in pursuit of reflection, etc. And more money can be spent on this than on the actual training of neural networks.
ps about deterministic algorithms, if it is possible to simplify the life of the neural network, by some other algorithm highlighting, for example, the angle of rotation of the image, rotating it always correctly, then the training sample can no longer be supplied with all the rotations, easing the load on the learning stage itself