Models (dosma.models)

OAI 2D U-Net

A 2D U-Net trained on a downsampled rendition of the OAI iMorphics DESS dataset :cite:`chaudhari2018open`. Inputs are zero-mean, unit standard deviation normalized before segmentation.

Aliases: oai-unet2d, oai_unet2d

IWOAI Segmentation Challenge - Team 6 2D U-Net

This model was submitted by Team 6 to the 2019 International Workshop on Osteoarthritis Segmentation :cite:`desai2020international`. It consists of a 2D U-Net trained on the standardized OAI training dataset.

Note, inputs are not normalized before segmentation and therefore may be difficult to generalize to DESS scans with different parameters than the OAI.

Aliases: iwoai-2019-t6

IWOAI Segmentation Challenge - Team 6 2D U-Net (Normalized)

This model is a duplicate of the iwoai-2019-t6 network (above), but differs in that it uses zero-mean, unit standard deviation normalized inputs. This may make the network more robust to different DESS scan parameters and/or scanner vendors.

While this model was not submitted to the IWOAI challenge, the architecture, training parameters, and dataset are identical to the Team 6 submission. Performance on the standardized OAI test set was similar to the original network submitted by Team 6 (see table below).

Aliases: iwoai-2019-t6-normalized

SKM-TEA qDESS Knee Segmentation - 2D U-net

This collection of models are trained on the SKM-TEA dataset (previously known as the 2021 Stanford qDESS Knee Dataset). Details of the different models that are trained are shown in the training configurations distributed with the weights.

• qDESS_2021_v1-rms-unet2d-pc_fc_tc_men_weights.h5: This is the baseline RSS model trained on the SKM-TEA v1 dataset. Though the same hyperparameters were used, this model (trained with Tensorflow/Keras) performs better than the PyTorch implementation specified in the main paper. Results are shown in the table below.
• qDESS_2021_v0_0_1-rms-pc_fc_tc_men_weights.h5: This model is trained on the 2021 Stanford qDESS knee dataset (v0.0.1).
• qDESS_2021_v0_0_1-traintest-rms-pc_fc_tc_men_weights.h5: This model is trained on both the train and test set of the 2021 Stanford qDESS knee dataset (v0.0.1).

Aliases: stanford-qdess-2021-unet2d, skm-tea-unet2d