Advanced Tips

Dicom Specialized Training (SSDL)

Smart Scan Deep Learning (SSDL)

Smart Scan Deep Learning (SSDL) is a advanced technique that can optimize the training results of Dicom images. SSDL will automatically activate when Dicom format in the dataset are detected.

Prevent Data Leakage

Data Leakage during training happens when the train/validate split is not handled carefully, causing some of the validation data contains information from the training data. As a result the validation score might overperform and be misleading.

In medical scans with image series, information might be shared between slices in a single scan. Therefore, splitting data by study prevents data leakage, as shown in the figure below.

SSDL prevents data leakage in sequential images through "split by study"

Activating SSDL

When the user selects a dataset in training tasks, the system will detect whether the dataset contains sequential information.

DeepQ AI Platform detects sequential image in a dataset and activates SSDL.

DeepQ AI Platform detects sequential image in a dataset and activates SSDL.

SSDL showing up in training task detail

AI Assisted Annotation

An algorithm trained to classify Chest X-ray images cannot help in segmenting tumors in brain MRI since each Deep learning algorithm is task-oriented. AI assisted annotation must start with an neural network trained specifically for the same purpose.

AI Training

• Create an prediction only batch inference.

• Select the dataset from the inference list, and click on the "EXPORT TO ANNOTATE" tab to start the AI assisted annotation task.

• Name the annotation data that you want to export.

• The column below shows the dataset details, where you can make sure if you have select the right one.

DeepCap

• Back to DeepCap, and create an annotation project.

• Select the dataset from the drop list, enable the option "Include annotation data", and choose the one that have been annotated by AI.