--- a
+++ b/rocaseg/datasets/dataset_okoa.py
@@ -0,0 +1,196 @@
+import os
+import glob
+import logging
+from collections import defaultdict
+
+import numpy as np
+import pandas as pd
+from joblib import Parallel, delayed
+
+from tqdm import tqdm
+import cv2
+from torch.utils.data.dataset import Dataset
+
+from rocaseg.datasets.constants import locations_f43h
+
+
+logging.basicConfig()
+logger = logging.getLogger('dataset')
+logger.setLevel(logging.DEBUG)
+
+
+def index_from_path_okoa(path_root, force=False):
+    fname_meta_dyn = os.path.join(path_root, 'meta_dynamic.csv')
+    fname_meta_base = os.path.join(path_root, 'meta_base.csv')
+
+    if not os.path.exists(fname_meta_dyn) or force:
+        fnames_image = glob.glob(
+            os.path.join(path_root, '**', 'images', '*.png'), recursive=True)
+        logger.info('{} images found'.format(len(fnames_image)))
+        fnames_mask = glob.glob(
+            os.path.join(path_root, '**', 'masks', '*.png'), recursive=True)
+        logger.info('{} masks found'.format(len(fnames_mask)))
+
+        df_meta = pd.read_csv(fname_meta_base,
+                              dtype={'subset': str,
+                                     'patient': str,
+                                     'release': str,
+                                     'sequence': str,
+                                     'side': str,
+                                     'slice_idx': int,
+                                     'pixel_spacing_0': float,
+                                     'pixel_spacing_1': float,
+                                     'slice_thickness': float,
+                                     'KL': int},
+                              index_col=False)
+
+        if len(fnames_image) != len(df_meta):
+            raise ValueError("Number of images doesn't match with the metadata")
+        if len(fnames_mask) != len(df_meta):
+            raise ValueError("Number of masks doesn't match with the metadata")
+
+        df_meta['path_image'] = [os.path.join(path_root, e)
+                                 for e in df_meta['path_rel_image']]
+        df_meta['path_mask'] = [os.path.join(path_root, e)
+                                for e in df_meta['path_rel_mask']]
+
+        # Check for in-slice mask presence
+        def worker_74n(fname):
+            mask = read_mask(path_file=fname, mask_mode='raw')
+            if np.any(mask[1:] > 0):
+                return 1
+            else:
+                return 0
+
+        logger.info('Exploring the annotations')
+        tmp = Parallel(n_jobs=-1)(
+            delayed(worker_74n)(row['path_mask'])
+            for _, row in tqdm(df_meta.iterrows(), total=len(df_meta)))
+        df_meta['has_mask'] = tmp
+
+        # Sort the records
+        df_meta_sorted = (df_meta
+                          .sort_values(['patient', 'sequence', 'slice_idx'])
+                          .reset_index()
+                          .drop('index', axis=1))
+
+        df_meta_sorted.to_csv(fname_meta_dyn, index=False)
+    else:
+        df_meta_sorted = pd.read_csv(fname_meta_dyn,
+                                     dtype={'subset': str,
+                                            'patient': str,
+                                            'release': str,
+                                            'sequence': str,
+                                            'side': str,
+                                            'slice_idx': int,
+                                            'pixel_spacing_0': float,
+                                            'pixel_spacing_1': float,
+                                            'slice_thickness': float,
+                                            'KL': int,
+                                            'has_mask': int},
+                                     index_col=False)
+    return df_meta_sorted
+
+
+def read_image(path_file):
+    image = cv2.imread(path_file, cv2.IMREAD_GRAYSCALE)
+    return image.reshape((1, *image.shape))
+
+
+def read_mask(path_file, mask_mode):
+    """Read mask from the file, and pre-process it.
+
+    IMPORTANT: currently, we handle the inter-class collisions by assigning
+    the joint pixels to a class with a lower index.
+
+    Parameters
+    ----------
+    path_file: str
+        Full path to mask file.
+    mask_mode: str
+        Specifies which channels of mask to use.
+
+    Returns
+    -------
+    out : (ch, d0, d1) uint8 ndarray
+
+    """
+    mask = cv2.imread(path_file, cv2.IMREAD_GRAYSCALE)
+
+    locations = {
+        '_background': (locations_f43h['_background'],),
+        'femoral': (locations_f43h['femoral'],),
+        'tibial': (locations_f43h['tibial'], ),
+    }
+
+    if mask_mode == 'raw':
+        return mask
+    elif mask_mode == 'background_femoral_unitibial':
+        ret = np.empty((3, *mask.shape), dtype=mask.dtype)
+        ret[0, :, :] = np.isin(mask, locations['_background']).astype(np.uint8)
+        ret[1, :, :] = np.isin(mask, locations['femoral']).astype(np.uint8)
+        ret[2, :, :] = np.isin(mask, locations['tibial']).astype(np.uint8)
+        return ret
+    else:
+        raise ValueError('Invalid `mask_mode`')
+
+
+class DatasetOKOASagittal2d(Dataset):
+    def __init__(self, df_meta, mask_mode=None, name=None, transforms=None,
+                 sample_mode='x_y', **kwargs):
+        logger.warning('Redundant dataset init arguments:\n{}'
+                       .format(repr(kwargs)))
+
+        self.df_meta = df_meta
+        self.mask_mode = mask_mode
+        self.name = name
+        self.transforms = transforms
+        self.sample_mode = sample_mode
+
+    def __len__(self):
+        return len(self.df_meta)
+
+    def _getitem_x_y(self, idx):
+        image = read_image(self.df_meta['path_image'].iloc[idx])
+        mask = read_mask(self.df_meta['path_mask'].iloc[idx], self.mask_mode)
+
+        # Apply transformations
+        if self.transforms is not None:
+            for t in self.transforms:
+                if hasattr(t, 'randomize'):
+                    t.randomize()
+                image, mask = t(image, mask)
+
+        tmp = dict(self.df_meta.iloc[idx])
+        tmp['image'] = image
+        tmp['mask'] = mask
+
+        tmp['xs'] = tmp['image']
+        tmp['ys'] = tmp['mask']
+        return tmp
+
+    def __getitem__(self, idx):
+        if self.sample_mode == 'x_y':
+            return self._getitem_x_y(idx)
+        else:
+            raise ValueError('Invalid `sample_mode`')
+
+    def read_image(self, path_file):
+        return read_image(path_file=path_file)
+
+    def read_mask(self, path_file):
+        return read_mask(path_file=path_file, mask_mode=self.mask_mode)
+
+    def describe(self):
+        summary = defaultdict(float)
+        for i in range(len(self)):
+            if self.sample_mode == 'x_y':
+                _, mask = self.__getitem__(i)
+            else:
+                mask = self.__getitem__(i)['mask']
+            summary['num_class_pixels'] += mask.numpy().sum(axis=(1, 2))
+        summary['class_importance'] = \
+            np.sum(summary['num_class_pixels']) / summary['num_class_pixels']
+        summary['class_importance'] /= np.sum(summary['class_importance'])
+        logger.info('Dataset statistics:')
+        logger.info(sorted(summary.items()))