It’s time to speed up Deep Learning so it can digest any data like never before

Deep Learning should be also fun and fast

Deep learning is now massively applied to healthcare diagnostics (Oxford University Press) and medical AI is on all investment radars, constantly looking for those who stand out of the crowd.

To be the best you are chasing quality and speed of product delivery, and while DL algorithms are well known — it is still a struggle to get state of the art (SOTA) results “like in papers”, not to mention create something better and faster.

Today, you will know how to get full speed of applied algorithms, with no matter what data runs through it:

step 0: Try ready to use ML models and proven NN architectures on any data — uber easy and fast. More models are being loaded for public regularly.

step 1: Build easy and clean pipelines with no prior experience. As for example, if you want to build NN for medical predictions here is a sample of a pipeline that loads cardio signals, makes preprocessing and trains a model for 50 epochs in just 15 lines :

Basic pipeline example

Take your knowledge to the new level by assembling flexible pipelines that ease the training and development process even in complex data, as CT scans:

Sample of advanced pipeline for NN on medical dataset of DICOM images

step 2: Speed up annoyingly time-consuming preprocessing operations of data-batches with "inbatch_parallelisation":

numba.njit + dataset.inbatch_parallel = performance boost

What use does it make for people?

Without the influence of medical universities — none of this would make much sense and most likely would stay unnoticed.
But, after healthcare datasets started to become public, here is short list of tasks that you apply your new knowledge to:

1. Heart disease recognition by ECG . Physionet databank expands annually
2. Breast Cancer segmentatition and processing of digitised images of a fine needle aspirate (FNA). Also, large number of marked up DICOM can be accessed in Cancer Imaging Archive
3. Lung Cancer detection with Computer tomography (DICOM)
4. Brain activity classification on EEG waves
5. NHS x-rays and MRI scans analysis from Diagnostic Imaging Dataset (DID)