This repository contains the code for the first place in the competition on Low-dose Computed Tomography Perceptual Image Quality Assessment. The data needed to train all models can be downloaded from zenodo. The structure of this project is relatively straightforward, once the data is in place. A jupyter notebook shows how I split the data into five train/validation sets. As a result, 10 csv files are generated, e.g. tr_f1.csv, vl_f1.csv. These can then be used for carrying out a training run, and we do five of them per model, one is a CNN (resnext50) and the other a transformer (swin), like this:
python train_PMH.py --csv_train data/train_f1.csv --model swin --n_heads 4 --overall_loss rps --hypar 10 --save_path swin/4hml_rps10_f1_104 --seed 1 --cycle_lens 10/4
python train_PMH.py --csv_train data/train_f2.csv --model swin --n_heads 4 --overall_loss rps --hypar 10 --save_path swin/4hml_rps10_f2_104 --seed 2 --cycle_lens 10/4
python train_PMH.py --csv_train data/train_f3.csv --model swin --n_heads 4 --overall_loss rps --hypar 10 --save_path swin/4hml_rps10_f3_104 --seed 3 --cycle_lens 10/4
python train_PMH.py --csv_train data/train_f4.csv --model swin --n_heads 4 --overall_loss rps --hypar 10 --save_path swin/4hml_rps10_f4_104 --seed 4 --cycle_lens 10/4
python train_PMH.py --csv_train data/train_f5.csv --model swin --n_heads 4 --overall_loss rps --hypar 10 --save_path swin/4hml_rps10_f5_104 --seed 5 --cycle_lens 10/4
Probably the most interesting bit here is that we use multi-head models, which we have shown to be nicely calibrated:
Multi-Head Multi-Loss Model Calibration
Adrian Galdran, Johan Verjans, Gustavo Carneiro, and Miguel A. González Ballester
MICCAI 2024 [link](https://arxiv.org/abs/2303.01099)
Also, the loss function (which you can find in the utils folder) is inspired in the Ranked Probabilty Score idea, which we reviewed here:
Performance Metrics for Probabilistic Ordinal Classifiers
Adrian Galdran, MICCAI 2024 [link](https://arxiv.org/abs/2309.08701)
A separate jupyter notebook illustrate how the RPS loss works.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent