Consider Gaussian function for single pixel annotation

[jcohenadad]

Currently, a window of 50 samples, centered around the epileptic peak, is labeled with value 1.

The problem with that, is that the model is exposed to values far away from the peak, with values 1.

A suggestion would be to, instead of using a step function, to use a Gaussian or Gamma function, centered around the peak.

We would use the SoftSeg training losses.

[mpompolas]

Adjusted the brainstorm code to take that into account, with a checkbox for activating it:

The Gaussian weighting is applied on the time dimension only.

Just started training using this configuration file with SoftSeg based on a discussion I had with @charleygros in the past, but the discussion was about 2D, not 3D training.
@naga-karthik please take a look at it when you get a chance.

Config file

{
  "command": "train",
  "gpu_ids": [
    0
  ],
  "path_output": "/home/GRAMES.POLYMTL.CA/u111358/data_nvme_u111358/EEG-ivado/epilepsy/distant_spikes_2_seconds_gaussian_annotation/output_Epilepsy_EEG_runs_3_gaussian_annotation1",
  "model_name": "EEG_model",
  "debugging": true,
  "object_detection_params": {
    "object_detection_path": null,
    "safety_factor": [
      1,
      1,
      1
    ]
  },
  "loader_parameters": {
    "path_data": "/home/GRAMES.POLYMTL.CA/u111358/data_nvme_u111358/EEG-ivado/epilepsy/distant_spikes_2_seconds_gaussian_annotation/data_Epilepsy_EEG_runs_3_gaussian_annotation",
    "subject_selection": {
      "n": [],
      "metadata": [],
      "value": []
    },
    "target_suffix": [
      "_saw_EST_distant"
    ],
    "roi_params": {
      "suffix": null,
      "slice_filter_roi": null
    },
    "contrast_params": {
      "training_validation": [
        "EEGarounddistantspikes"
      ],
      "testing": [
        "EEGarounddistantspikes"
      ],
      "balance": {}
    },
    "slice_filter_params": {
      "filter_empty_mask": false,
      "filter_empty_input": true
    },
    "slice_axis": "sagittal",
    "multichannel": false,
    "soft_gt": true
  },
  "split_dataset": {
    "fname_split": null,
    "random_seed": 6,
    "center_test": [],
    "method": "per_patient",
    "balance": null,
    "train_fraction": 0.6,
    "test_fraction": 0.2
  },
  "training_parameters": {
    "batch_size": 8,
    "loss": {
      "name": "AdapWingLoss", 
      "theta": 0.5, 
      "alpha": 2.1, 
      "epsilon": 1, 
      "omega": 8
    },
    "training_time": {
      "num_epochs": 100,
      "early_stopping_patience": 50,
      "early_stopping_epsilon": 0.001
    },
    "scheduler": {
      "initial_lr": 0.001,
      "lr_scheduler": {
        "name": "CosineAnnealingLR",
        "base_lr": 1E-5,
        "max_lr": 0.01
      }
    },
    "balance_samples": {
      "applied": false,
      "type": "gt"
    },
    "mixup_alpha": null,
    "transfer_learning": {
      "retrain_model": null,
      "retrain_fraction": 1,
      "reset": true
    }
  },
  "default_model": {
    "name": "Unet",
    "dropout_rate": 0.3,
    "bn_momentum": 0.9,
    "depth": 3,
    "final_activation": "relu"
  },
  "FiLMedUnet": {
    "applied": false,
    "metadata": "institution_id",
    "film_layers": [
      0,
      1,
      0,
      0,
      0,
      0,
      0,
      0
    ]
  },
  "Modified3DUNet": {
    "applied": true,
    "length_3D": [
      112,
      112,
      128
    ],
    "stride_3D": [
      112,
      112,
      64
    ],
    "attention": false,
    "n_filters": 8
  },
  "uncertainty": {
    "epistemic": false,
    "aleatoric": false,
    "n_it": 0
  },
  "postprocessing": {
    "remove_noise": {
      "thr": -1
    },
    "binarize_prediction": {
      "thr": 0.5
    },
    "uncertainty": {
      "thr": -1,
      "suffix": "_unc-vox.nii.gz"
    },
    "fill_holes": {},
    "remove_small": {
      "unit": "vox",
      "thr": 3
    }
  },
  "evaluation_parameters": {
    "target_size": {
      "unit": "vox",
      "thr": [
        20,
        100
      ]
    },
    "overlap": {
      "unit": "vox",
      "thr": 3
    }
  },
  "transformation": {
    "CenterCrop": {
      "size": [
        112,
        112,
        512
      ]
    },
    "RandomAffine": {
      "degrees": 0,
      "translate": [
        0,
        0,
        0.03
      ],
      "applied_to": [
        "im",
        "gt"
      ],
      "dataset_type": [
        "training"
      ]
    },
    "NumpyToTensor": {},
    "NormalizeInstance": {
      "applied_to": [
        "im"
      ]
    }
  },
  "brainstorm": {
    "modality": "EEG",
    "event_for_ground_truth": "saw_EST_distant",
    "annotations_time_window": [
      -0.1,
      0.3
    ],
    "resampling": 100,
    "jitter": 0,
    "soft_annotation_threshold": [],
    "bids_folder_creation_mode": "Separate each trial as different subjects",
    "sInputs": [
      "1: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial001_notch_band_resample_time.mat",
      "2: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial002_notch_band_resample_time.mat",
      "3: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial003_notch_band_resample_time.mat",
      "4: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial006_notch_band_resample_time.mat",
      "5: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial007_notch_band_resample_time.mat",
      "6: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial008_notch_band_resample_time.mat",
      "7: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial009_notch_band_resample_time.mat",
      "8: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial011_notch_band_resample_time.mat",
      "9: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial015_notch_band_resample_time.mat",
      "10: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial017_notch_band_resample_time.mat",
      "11: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial019_notch_band_resample_time.mat",
      "12: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial020_notch_band_resample_time.mat",
      "13: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial022_notch_band_resample_time.mat",
      "14: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial023_notch_band_resample_time.mat",
      "15: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial025_notch_band_resample_time.mat",
      "16: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial027_notch_band_resample_time.mat",
      "17: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial030_notch_band_resample_time.mat",
      "18: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial031_notch_band_resample_time.mat",
      "19: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial032_notch_band_resample_time.mat",
      "20: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial033_notch_band_resample_time.mat",
      "21: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial035_notch_band_resample_time.mat",
      "22: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial036_notch_band_resample_time.mat",
      "23: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial037_notch_band_resample_time.mat",
      "24: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial038_notch_band_resample_time.mat",
      "25: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial039_notch_band_resample_time.mat",
      "26: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial042_notch_band_resample_time.mat",
      "27: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial044_notch_band_resample_time.mat",
      "28: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial049_notch_band_resample_time.mat",
      "29: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial050_notch_band_resample_time.mat",
      "30: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial051_notch_band_resample_time.mat",
      "31: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial052_notch_band_resample_time.mat",
      "32: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial053_notch_band_resample_time.mat",
      "33: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial056_notch_band_resample_time.mat",
      "34: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial057_notch_band_resample_time.mat",
      "35: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial058_notch_band_resample_time.mat",
      "36: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial059_notch_band_resample_time.mat",
      "37: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial060_notch_band_resample_time.mat",
      "38: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial062_notch_band_resample_time.mat",
      "39: sub-pt0090/sub-pt0090_ses-20210722_task-rest_run-03_meg/data_around_distant_spikes_trial063_notch_band_resample_time.mat"
    ]
  }
}

[mpompolas]

First result is not very promising. Need to test with an even larger Kernel

Results

The slice_axis parameter was wrong in this test. Check 12

IGNORE RESULT

[jcohenadad]

First result is not very promising. Need to test with an even larger Kernel

Are we sure that trilinear interpolation without binarization is used for data augmentation? Is this information supposed to be present in the config file? @andreanne-lemay @charleygros