Merge pull request #2 from tomasprinda/ocontours_load

Loading ocontours added

README.md

@@ -29,8 +29,9 @@ python scripts/train.py --train_dir DATA_DIR/datasets/train/ --dev_dir DATA_DIR/
 
 which starts training of the net and it stores experiment data to [experiments/exp01/](experiments/exp01/) folder.
 
+## Phase 01
 
-## Answers to questions
+### Answers to questions
 
 How did you verify that you are parsing the contours correctly?
  - First I converted contours to mask by untested method for some random images 
@@ -63,7 +64,15 @@ Given the pipeline you have built, can you see any deficiencies that you would c
  - Automatic hyperparameter tuning
  - Better logging
 
-## Notes
+### Notes
  - The aim of this project was to prepare data for training and create training pipeline. Training phase has not been tuned. 
  More images would probably be neccessary.
 
+## Phase 2
+
+### Answers to questions
+
+After building the pipeline, please discuss any changes that you made to the pipeline you built in Phase 1, and why you made those changes.
+
+
+

experiments/test_imask_inside_omask_without_skipping/wrong_files.csv

@@ -0,0 +1,9 @@
+1;0.0068206787109375;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/119.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0119-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0119-ocontour-manual.txt
+2;0.0070953369140625;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/139.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0139-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0139-ocontour-manual.txt
+3;0.0076446533203125;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/159.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0159-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0159-ocontour-manual.txt
+4;0.0059356689453125;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/179.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0179-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0179-ocontour-manual.txt
+5;0.0039825439453125;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/199.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0199-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0199-ocontour-manual.txt
+6;0.003265380859375;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/219.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0219-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0219-ocontour-manual.txt
+7;0.007568359375;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/59.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0059-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0059-ocontour-manual.txt
+8;0.0051422119140625;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/79.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0079-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0079-ocontour-manual.txt
+9;0.0067596435546875;/www/data/prinda/ventricle_segmentation/final_data/dicoms/SCD0000501/99.dcm;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/i-contours/IM-0001-0099-icontour-manual.txt;/www/data/prinda/ventricle_segmentation/final_data/contourfiles/SC-HF-I-6/o-contours/IM-0001-0099-ocontour-manual.txt

tests/test_parsing.py

@@ -1,10 +1,16 @@
 import unittest
 
 import os
+from pprint import pprint
+
+import matplotlib
+matplotlib.use('Agg')  # prevents _tkinter error
+import matplotlib.pyplot as plt
 
 from ventricle_segmentation import cfg
-from ventricle_segmentation.parsing import parse_dicom_file, np, polygon_to_mask, parse_contour_file, load_all_scans
-from ventricle_segmentation.utils import print_info, pickle_load, iou
+from ventricle_segmentation.core import AnnotatedScan, ContourMask
+from ventricle_segmentation.parsing import parse_dicom_file, np, polygon_to_mask, parse_contour_file, load_all_scans, get_contour_mask
+from ventricle_segmentation.utils import print_info, pickle_load, iou, prepare_exp_dir, csv_dump, plot_annotated_scan
 
 
 class TestParsing(unittest.TestCase):
@@ -24,30 +30,89 @@ def test_parse_dicom_file(self):
             dicom_img = parse_dicom_file(dicom_file)
             self.assertIsInstance(dicom_img, np.ndarray)
 
-    def test_poly_to_mask(self):
-        for test_mask_file in os.listdir(cfg.TEST_MASKS_DIR):
+    def test_get_contour_mask(self):
+        test_files = list(os.listdir(cfg.TEST_MASKS_DIR))
+
+        # Test if some test images
+        self.assertGreater(len(test_files), 0)
+
+        for test_mask_file in test_files:
             test_mask_file = os.path.join(cfg.TEST_MASKS_DIR, test_mask_file)
-            expected_annotated_scan = pickle_load(test_mask_file)
+            expected_annotated_scan = pickle_load(test_mask_file)  # type: AnnotatedScan
+
+            # Test masks
+            masks_dict = {"imask": expected_annotated_scan.imask, "omask": expected_annotated_scan.omask}
+            for mask_type, expected_mask in masks_dict.items():
+                if expected_mask is not None:
+                    shape = expected_annotated_scan.dicom_img.shape
+                    imask = get_contour_mask(expected_mask.contours_file, shape)
+
+                    self.assertEqual(imask.mask.tolist(), expected_mask.mask.tolist(), mask_type)
+                    self.assertEqual(imask.contours_file, expected_mask.contours_file, mask_type)
+                else:
+                    self.assertIsNone(expected_mask, mask_type)
+
+        non_existing_path = "/not/existing/file.txt"
+        mask = get_contour_mask(non_existing_path, (100, 100))
+        self.assertIsNone(mask)
+
+    def test_imask_inside_omask(self):
+        prepare_exp_dir("test_imask_inside_omask", clean_dir=True)
+        i_wrong = 0
 
-            expected_mask = expected_annotated_scan.imask
+        tolerance = 0.001
+        wrong_scores = []
+        for annotated_scan in load_all_scans(cfg.LINKS_FILE, cfg.DICOMS_DIR, cfg.CONTOURS_DIR):
 
-            contours = parse_contour_file(expected_annotated_scan.icontours_file)
-            mask = polygon_to_mask(contours, *expected_mask.shape)
+            if annotated_scan.imask is None or annotated_scan.omask is None:
+                continue  # Skip if some mask not available
 
-            result = iou(mask, expected_mask)
-            expected_result = 1
-            tolerance = 0.05
+            wrong_pixels = annotated_scan.imask.mask > annotated_scan.omask.mask
+            wrong_pixels = np.mean(np.ravel(wrong_pixels))
 
-            self.assertLess(abs(result - expected_result), tolerance)
+            # Plot incorrect data
+            if wrong_pixels > tolerance:
+                i_wrong += 1
+                row = [i_wrong, wrong_pixels, annotated_scan.dicom_file, annotated_scan.imask.contours_file, annotated_scan.omask.contours_file]
+                csv_dump([row], os.path.join(cfg.EXP_DIR, "wrong_files.csv"), append=True)
+                plot_annotated_scan(annotated_scan)
+                plt.savefig(os.path.join(cfg.EXP_DIR, "{}.png".format(i_wrong)))
+                plt.clf()
+
+            wrong_scores.append(wrong_pixels)
+
+        # Test it
+        for wrong_pixels in wrong_scores:
+            self.assertLess(wrong_pixels, tolerance)
 
     def test_load_all_scans(self):
-        annotated_scans = list(load_all_scans(cfg.LINKS_FILE, cfg.DICOMS_DIR, cfg.CONTOURS_DIR, n=20))
+        n = 20
+        annotated_scans = list(load_all_scans(cfg.LINKS_FILE, cfg.DICOMS_DIR, cfg.CONTOURS_DIR, n=n))
+
+        # Test if some test images
+        self.assertEqual(len(annotated_scans), n)
+
         for annotated_scan in annotated_scans:
-            self.assertIsInstance(annotated_scan.imask, np.ndarray)
-            self.assertIsInstance(annotated_scan.dicom_img, np.ndarray)
-            self.assertTupleEqual(annotated_scan.dicom_img.shape, annotated_scan.imask.shape)
 
+            imask_loaded = isinstance(annotated_scan.imask, ContourMask)
+            omask_loaded = isinstance(annotated_scan.omask, ContourMask)
+            self.assertTrue(imask_loaded or omask_loaded)
+
+            self.assertIsInstance(annotated_scan.dicom_img, np.ndarray)
 
+            if imask_loaded:
+                self.assertTupleEqual(annotated_scan.dicom_img.shape, annotated_scan.imask.mask.shape)
+                self.assertIsInstance(annotated_scan.imask.contours_file, str)
+                self.assertNotEqual(annotated_scan.imask.contours_file, "")
+            else:
+                self.assertIsNone(annotated_scan.imask)
+
+            if omask_loaded:
+                self.assertTupleEqual(annotated_scan.dicom_img.shape, annotated_scan.omask.mask.shape)
+                self.assertIsInstance(annotated_scan.omask.contours_file, str)
+                self.assertNotEqual(annotated_scan.omask.contours_file, "")
+            else:
+                self.assertIsNone(annotated_scan.omask)
 
 
 if __name__ == '__main__':

ventricle_segmentation/core.py

@@ -3,14 +3,28 @@ class AnnotatedScan:
     Dataset example holder
     """
 
-    def __init__(self, dicom_img, imask, dicom_file, icontours_file):
+    def __init__(self, dicom_img, dicom_file, imask, omask):
         """
         :param np.ndarray dicom_img:
-        :param np.ndarray imask: dtype=bool Mask that defines left ventricular blood pool in dicom image
         :param str dicom_file: Path to dicom file
-        :param str icontours_file: Path to countours file (before converting to mask)
+        :param ContourMask|None imask: Mask that separates the left ventricular blood pool from the heart muscle (myocardium)
+        :param ContourMask|None omask: Mask that defines the outer border of the left ventricular heart muscle
         """
         self.dicom_img = dicom_img
-        self.imask = imask
         self.dicom_file = dicom_file
-        self.icontours_file = icontours_file
+        self.imask = imask
+        self.omask = omask
+
+
+class ContourMask:
+    """
+    i-contour / o-contour mask holder
+    """
+
+    def __init__(self, mask, contours_file):
+        """
+        :param np.ndarray mask: dtype=bool i/o-contour boolean mask
+        :param str contours_file: Path to i/o-countours file (before converting to mask)
+        """
+        self.mask = mask
+        self.contours_file = contours_file

ventricle_segmentation/dataset_loader.py

@@ -29,9 +29,9 @@ def __getitem__(self, index):
         dicom_img = (annotated_scan.dicom_img.astype(np.float32) - 500) / 1000
         dicom_img = np.expand_dims(dicom_img, 0)  # adding channel at dimension 0, required dims in batch (N,C_in,H_in,W_in); N is not present yet
 
-        imask = annotated_scan.imask.astype(np.int64)
+        imask = annotated_scan.imask.mask.astype(np.int64)
 
-        return dicom_img, imask, annotated_scan.dicom_file, annotated_scan.icontours_file
+        return dicom_img, imask, annotated_scan.dicom_file, annotated_scan.imask.contours_file
 
     def __len__(self):
         return len(self.scan_pkl_files)

ventricle_segmentation/parsing.py

@@ -8,7 +8,7 @@
 from PIL import Image, ImageDraw
 from pydicom.errors import InvalidDicomError
 
-from ventricle_segmentation.core import AnnotatedScan
+from ventricle_segmentation.core import AnnotatedScan, ContourMask
 
 
 def load_all_scans(links_file, dicoms_dir, contours_dir, n=None):
@@ -24,10 +24,16 @@ def load_all_scans(links_file, dicoms_dir, contours_dir, n=None):
     i = 0
     for (dicom_id, contours_id) in parse_links_file(links_file):
 
+        if dicom_id == "SCD0000501":
+            # o-contours failed TestParsing.test_imask_inside_omask().
+            # See experiments/test_imask_inside_omask_without_skipping
+            continue
+
         dicom_patients_dir = os.path.join(dicoms_dir, dicom_id)
-        contour_patients_dir = os.path.join(contours_dir, contours_id, "i-contours")
+        icontour_patients_dir = os.path.join(contours_dir, contours_id, "i-contours")
+        ocontour_patients_dir = os.path.join(contours_dir, contours_id, "o-contours")
 
-        for annotated_scan in load_patient_scans(dicom_patients_dir, contour_patients_dir):
+        for annotated_scan in load_patient_scans(dicom_patients_dir, icontour_patients_dir, ocontour_patients_dir):
 
             if n is not None and i >= n:
                 return
@@ -53,32 +59,69 @@ def parse_links_file(filename):
     return links_lst
 
 
-def load_patient_scans(dicom_patient_dir, icontour_patient_dir):
+pattern_dicom = re.compile("^(\d+)\.dcm")
+
+
+def load_patient_scans(dicom_patient_dir, icontour_patient_dir, ocontour_patient_dir):
     """
-    Loads contours from icontour_patient_dir and loads also corresponding dicom images for annotations from dicom_patient_dir.
+    Loads dicom files with corresponding i-contours and o-contours files if at least one of the contours exists
+    for the dicom file.
     Make sure dicom_patient_dir matches icontour_patient_dir (by link.csv)
 
     :param str dicom_patient_dir: dir containing dicom files {:d}.dcm files
-    :param icontour_patient_dir: dir contating contour files IM-0001-{:04d}-icontour-manual.txt
+    :param str icontour_patient_dir: dir contating i-contour files IM-0001-{:04d}-icontour-manual.txt
+    :param str ocontour_patient_dir: dir contating o-contour files IM-0001-{:04d}-ocontour-manual.txt
     :return list[AnnotatedScan]: iterable
     """
-    pattern = re.compile("^IM-0001-(\d+)-icontour-manual\.txt")
-    for icontour_file in os.listdir(icontour_patient_dir):
-        match = pattern.match(icontour_file)
-        icontour_file = os.path.join(icontour_patient_dir, icontour_file)
+
+    for dicom_file in os.listdir(dicom_patient_dir):
+        match = pattern_dicom.match(dicom_file)
 
         if match:  # Correct icontour file
+
             dicom_nr = int(match.group(1))
-            dicom_file = os.path.join(dicom_patient_dir, "{}.dcm".format(dicom_nr))
-            dicom_img = parse_dicom_file(dicom_file)
+            dicom_file = os.path.join(dicom_patient_dir, dicom_file)
+            icontour_file = os.path.join(icontour_patient_dir, "IM-0001-{:04d}-icontour-manual.txt".format(dicom_nr))
+            ocontour_file = os.path.join(ocontour_patient_dir, "IM-0001-{:04d}-ocontour-manual.txt".format(dicom_nr))
 
-            contours = parse_contour_file(icontour_file)
-            imask = polygon_to_mask(contours, *dicom_img.shape)
+            if not os.path.isfile(icontour_file) and not os.path.isfile(ocontour_file):
+                continue  # Don't load when no annotation data
 
-            annotated_scan = AnnotatedScan(dicom_img, imask, dicom_file, icontour_file)
+            annotated_scan = load_annotated_scan(dicom_file, icontour_file, ocontour_file)
             yield annotated_scan
 
 
+def load_annotated_scan(dicom_file, icontour_file, ocontour_file):
+    """
+    Loads AnnotatedScan from dicom and i/o-contours files.
+    If i/o-contours files deosn't exists, AnnotatedScan.imask/omask is None
+    :param str dicom_file: Path to dicom file
+    :param str icontour_file: Path to icontour file (might not exists)
+    :param str ocontour_file: Path to icontour file (might not exists)
+    :return AnnotatedScan :
+    """
+    dicom_img = parse_dicom_file(dicom_file)
+    imask = get_contour_mask(icontour_file, dicom_img.shape)
+    omask = get_contour_mask(ocontour_file, dicom_img.shape)
+
+    return AnnotatedScan(dicom_img, dicom_file, imask, omask)
+
+
+def get_contour_mask(contour_file, shape):
+    """
+    Loads contours file and converts it to mask
+    :param str contour_file: Contours file to load
+    :param (int, int) shape: Size of mask to create == size of a dicom file
+    :return ContourMask|None: Returns None if file doesn't exists
+    """
+    if not os.path.isfile(contour_file):
+        return None
+
+    contours = parse_contour_file(contour_file)
+    mask = polygon_to_mask(contours, *shape)
+    return ContourMask(mask, contour_file)
+
+
 def parse_contour_file(filename):
     """
     Parse the given contour filename
@@ -146,4 +189,3 @@ def polygon_to_mask(polygon, width, height):
     ImageDraw.Draw(img).polygon(xy=polygon, outline=0, fill=1)
     mask = np.array(img).astype(bool)
     return mask
-

ventricle_segmentation/utils.py

@@ -49,23 +49,33 @@ def plot_annotated_scan(annotated_scan, plot_mask=True):
     :param AnnotatedScan annotated_scan:
     """
     plt.figure(figsize=(12, 12))
-    plt.imshow(annotated_scan.dicom_img.astype(float))
+    plt.imshow(annotated_scan.dicom_img.astype(float), cmap=pl.cm.hot)
     if plot_mask:
-        plt.figure(figsize=(12, 12))
-        plt.imshow(annotated_scan.dicom_img.astype(float))
-        # Choose colormap
-        cmap = pl.cm.Reds
+        if annotated_scan.imask is not None:
+            plot_contour_mask(annotated_scan.dicom_img, annotated_scan.imask, cmap=pl.cm.Blues)
+        if annotated_scan.omask is not None:
+            plot_contour_mask(annotated_scan.dicom_img, annotated_scan.omask, cmap=pl.cm.Greens)
+    plt.title("\n".join([
+        annotated_scan.dicom_file,
+        annotated_scan.imask.contours_file if annotated_scan.imask else "no imask",
+        annotated_scan.omask.contours_file if annotated_scan.omask else "no omask"
+    ]))
 
-        # Get the colormap colors
-        my_cmap = cmap(np.arange(cmap.N))
 
-        # Set alpha
-        my_cmap[:, -1] = np.linspace(0, 1, cmap.N)
+def plot_contour_mask(dicom_img, contour_mask, cmap):
+    # plt.figure(figsize=(12, 12))
+    # plt.imshow(dicom_img.astype(float), cmap=pl.cm.hot)
 
-        # Create new colormap
-        my_cmap = ListedColormap(my_cmap)
+    # Get the colormap colors
+    my_cmap = cmap(np.arange(cmap.N))
 
-        plt.imshow(annotated_scan.imask.astype(float) / 4, cmap=my_cmap, alpha=.5)
+    # Set alpha
+    my_cmap[:, -1] = np.linspace(0, 1, cmap.N)
+
+    # Create new colormap
+    my_cmap = ListedColormap(my_cmap)
+
+    plt.imshow(contour_mask.mask.astype(float) / 4, cmap=my_cmap, alpha=.5)
 
 
 def prepare_exp_dir(exp_name, clean_dir):