2

Signed-off-by: myron <[email protected]>

Author: myron

auto3dseg/tasks/kits23/README.md

@@ -4,21 +4,20 @@
 TODO
 
 
-The KiTS dataset is from MICCAI 2023 challenge **[HEad and NeCK TumOR Segmentation and Outcome Prediction (HECKTOR22)](https://hecktor.grand-challenge.org)**. The solution described here won the 1st place in the HECKTOR22 challenge [(NVAUTO team)](https://hecktor.grand-challenge.org/final-leaderboard/):
+The KiTS dataset is from MICCAI 2023 challenge **[The 2023 Kidney and Kidney Tumor Segmentation Challenge (KiTS23)](https://kits-challenge.org/kits23/)**. The solution described here won the 1st place in the KiTS challenge [(NVAUTO team)](https://kits-challenge.org/kits23/#kits23-official-results):
 
-Andriy Myronenko, Md Mahfuzur Rahman Siddiquee, Dong Yang, Yufan He and Daguang Xu: "Automated head and neck tumor segmentation from 3D PET/CT". In MICCAI (2022). [arXiv](https://arxiv.org/abs/2209.10809)
+Andriy Myronenko, Dong Yang, Yufan He and Daguang Xu: "Automated 3D Segmentation of Kidneys and Tumors in MICCAI KiTS 2023 Challenge". In MICCAI (2023). [arXiv](https://arxiv.org/abs/2310.04110)
 
-![hecktor_PET_CT](./hecktor_data.jpg)
+![kits23_example](./kits23_example.png)
 
 ## Task overview
 
-The task is to segment 3D Head and Neck (H&N) tumors and lymph nodes classes from a pair of 3D CT and PET images.  The ground truth labels are provided for 524 cases with average 3D CT size of 512x512x200 voxels at 0.98x0.98x3 mm average resolution, and with average 3D PET size of 200x200x200 voxels at 4x4x4 mm. The CT and PET images where rigidly aligned to a common origin, but remain at different sizes and resolutions.
+The task is to segment kidneys, tumors and cysts from 3D CTs.  The ground truth labels are provided for 489 cases with resolutions between 0.39x0.39x0.5 and 1x1x5 mm. 
 
 
 ## Auto3DSeg
 
-The HECKTOR22 tutorial is only supported for **SegResNet** algo (since currently it is the only algo with support of multi-resolution input images, such as CT and PET).
-Auto3DSeg runs a full workflow including data analysis, and multi-fold training. Please download the dataset into /data/hecktor22 folder first.
+The KiTS tutorial is only supported for **SegResNet** algo, Auto3DSeg runs a full workflow including data analysis, and multi-fold training. Please download the dataset into /data/kits23 folder first.
 
 
 ### Running based on the input config
@@ -29,31 +28,27 @@ The Auto3DSeg can be run using a config **input.yaml**
 python -m monai.apps.auto3dseg AutoRunner run --input='./input.yaml' --algos='segresnet'
 ```
 
-### Running from python
 
-Alternatively you can also run Auto3DSeg from a python script, where you can customize more options. Please see the comments in **hecktor22.py**
-```bash
-python hecktor22.py
-```
-
-
-## Validation performance: NVIDIA DGX-1 (8x V100 16G)
+## Validation performance: NVIDIA DGX-1 (8x V100 32G)
 
-The validation results can be obtained by running the training script with MONAI 1.1.0 on NVIDIA DGX-1 with (8x V100 16GB) GPUs. The results below are in terms of **Aggregated Dice**, which was used as the key metric in the challenge [1,2]. The values of the Aggregated Dice slightly differ from a conventional average Dice (which is used by Auto3DSeg by default for all tasks).
+The validation results can be obtained by running the training script with MONAI 1.3.0 on NVIDIA DGX-1 with (8x V100 32GB) GPUs. The results below are in terms of average dice.
 
 
 | | Fold 0 | Fold 1 | Fold 2 | Fold 3 | Fold 4 | Avg |
 |:------:|:------:|:------:|:------:|:------:|:------:|:---:|
-| SegResNet | 0.7933 | 0.7862 | 0.7816 |0.8275 | 0.8059 | 0.7989 |
+| SegResNet | 0.8997 | 0.8739 | 0.8923 |0.8911 | 0.8892 |0.88924 |
 
 
 ## Data
 
-The HECKTOR22 challenge dataset [2,3] can be downloaded from [here](https://hecktor.grand-challenge.org) after the registration. Each user is responsible for checking the content of the datasets and the applicable licenses and determining if suitable for the intended use. The license for the HECKTOR22 dataset is different than MONAI license.
+The KiTS23 challenge dataset [2,3] can be downloaded from [here](https://kits-challenge.org/kits23). Each user is responsible for checking the content of the datasets and the applicable licenses and determining if suitable for the intended use. The license for the KiTS23 dataset is different than the MONAI license.
+
 
 ## References
-[1] Andriy Myronenko, Md Mahfuzur Rahman Siddiquee, Dong Yang, Yufan He and Daguang Xu: "Automated head and neck tumor segmentation from 3D PET/CT". In MICCAI (2022). https://arxiv.org/abs/2209.10809
+[1] Andriy Myronenko, Dong Yang, Yufan He and Daguang Xu: "Automated 3D Segmentation of Kidneys and Tumors in MICCAI KiTS 2023 Challenge". In MICCAI (2023). https://arxiv.org/abs/2310.04110
+
+
+[2] Heller, N., Isensee, F., Maier-Hein, K.H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., et al.: The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical Image Analysis 67, 101821 (2021)
 
-[2] Andrearczyk, V., Oreiller, V., Boughdad, S., Rest, C.C.L., Elhalawani, H., Jreige, M., Prior, J.O., Valli`eres, M., Visvikis, D., Hatt, M., Depeursinge, A.: Overview of the HECKTOR Challenge at MICCAI 2022: Automatic Head and Neck Tumor Segmentation and Outcome Prediction in PET/CT (2023), https://arxiv.org/abs/2201.04138
+[3] Heller, N., Wood, A., Isensee, F., Radsch, T., Tejpaul, R., Papanikolopoulos, N.,Weight, C.: The 2023 kidney and kidney tumor segmentation challenge, https://kits-challenge.org/kits23/
 
-[3] Oreiller, V., Andrearczyk, V., Jreige, M., Boughdad, S., Elhalawani, H., Castelli, J., Valli`eres, M., Zhu, S., Xie, J., Peng, Y., Iantsen, A., Hatt, M., Yuan, Y., Ma, J., Yang, X., Rao, C., Pai, S., Ghimire, K., Feng, X. Naser, M.A., Fuller, C.D., Yousefirizi, F., Rahmim, A., Chen, H., Wang, L., Prior, J.O., Depeursinge, A.: Head and neck tumor segmentation in PET/CT: The HECKTOR challenge. Medical Image Analysis 77, 102336 (2022)

auto3dseg/tasks/kits23/input.yaml

@@ -9,10 +9,20 @@ class_names:                        # names for tensorboard, and label index gro
   - { name: tumor, index: [2] }
 
 # use final sigmoid activation (instead of the default softmax), since KiTS regions are overlapping (multi-label segmentation)
-# this is optional to set, the system auto-detect overlapping labels.
+# this is optional to set, the system auto-detects overlapping labels automatically.
 sigmoid: true 
 
 
-roi_size: [192, 192, 192]
+# the config below is optional, but it explicitly sets params as it was used during KiTS23 challenge
+# otherwise, the defaults are (auto_scale_allowed is True) and the system will attempt to guess these settings according to the available GPU (e.g. make batch size larger) 
+auto_scale_allowed: false
+batch_size: 1
+roi_size: [256, 256, 256]
+num_epochs: 600
+resample: true
+resample_resolution: [0.78125, 0.78125, 0.78125]
+loss: {_target_: DiceLoss}
+
+