Function mergedpatch_gen

extract_features.py:496–536

(features, coords, dist_threshold=4, corr_threshold = 0.6)

    return [img, coords]

def mergedpatch_gen(features, coords, dist_threshold=4, corr_threshold = 0.6):

    # Get patch distance in pixels with rendered segmentation level. Note that each patch is squared and therefore same distance.
    patch_dist = abs(coords[0,2] - coords[0,0]) 
    print(patch_dist)
    
    # Compute feature similarity (cosine) and nearby pacthes (L2 norm - only need the top left x,y coordinates)
    cosine_matrix = cosine_similarity(features, features)
    coordinate_matrix = euclidean_distances(coords[:,:2], coords[:,:2])

    # NOTE: random selection for the first patch for patch merging might be less biased towards tissue orientation and size. 
    indices_avail = np.arange(features.shape[0])
    np.random.seed(0)  
    np.random.shuffle(indices_avail)

    # Merging together nearby patches and similar within pre-defined threshold. 
    mergedfeatures = []
    indices_used = []
    for ref in indices_avail:

        # This has been merged already
        if ref not in indices_used:

            # Making sure they won't be selected once more
            if indices_used:
                coordinate_matrix[ref,indices_used] = [np.Inf]*len(indices_used)
                cosine_matrix[ref,indices_used] = [0.0]*len(indices_used)
            
            indices_dist = np.where(coordinate_matrix[ref] < patch_dist*dist_threshold, 1 , 0)
            indices_corr = np.where(cosine_matrix[ref] > corr_threshold, 1 , 0)
            final_indices = indices_dist * indices_corr

            # which includes already the ref patch
            indices_used.extend(list(np.where(final_indices == 1)[0]))
            mergedfeatures.append(tuple((features[final_indices==1,:], coords[final_indices==1,:])))
        else:
            continue
        
    assert len(indices_used)==features.shape[0], f'Probably issue in contruscting merged features for graph {len(indices_used)}!={features.shape[0]}'

    return mergedfeatures

class HNSW:
    def __init__(self, space):