Page 280 - Computational Retinal Image Analysis

278 CHAPTER 14 OCT fluid detection and quantification

Lee et al. [22] used a U-net for a binary segmentation of IRF with a train-
ing and validation set composed of 1289 annotated B-scans. They achieved a
Dice similarity coefficient (DSC) of 0.73, close to the interobserver variability of
DSC = 0.75, on a test set of 30 B-scans, annotated by multiple experts. Roy et al.
[23] used a U-net to segment seven retinal layers and IRF jointly. They solved
the multiclass segmentation problem by combining Dice loss with a weighted lo-
gistic regression loss that compensated class imbalance and selectively penalized
misclassification. The model was successfully trained and tested on a small pub-
licly available dataset [24] containing 110 annotated B-scans from 10 patients
acquired with Spectralis OCT.
Two large validation studies recently showed that FCN can work effectively
across OCT device vendors and macular diseases. Venhuizen et al. [25] implemented
a multiscale network that used a range of contextual windows to segment IRF. The
method consisted of a cascade of two U-nets with two complementary tasks: the first
one aimed at delimiting the retinal region [26] and the second one at segmenting
IRF by integrating the output of the first U-net as both an additional input channel
and a constraining weight map used in computing the loss during the training. A
total of 221 OCT volumes from 151 patients (6158 B-scans) were used from which
the testing and evaluation had been performed on 99 OCT volumes (2487 B-scans)
from 75 patients. To obtain segmentations from scans of different OCT vendors, a
small amount of vendor-specific data was used for fine tuning. They reported a DSC
of 0.79 and, furthermore, demonstrated a good robustness and generalization even
without the fine tuning (DSC = 0.72).
Schlegl et al. [21] developed a semantic segmentation network to seg-
ment the retina, IRF, and SRF simultaneously (Fig. 3). The method was cross-
validated on a dataset of 354 fully annotated OCT volumes comprising three
main exudative diseases, such as nAMD (212), DME (32), and RVO (110), and
two OCT vendors, such as Cirrus (268) and Spectralis (86). Models for the two
OCT vendors were trained separately as it was shown that the Cirrus model
generalized well to a Spectralis dataset but not vice versa. Due to a difference
in fluid distributions between diseases, the model was first trained on scans
from patients with nAMD and RVO, followed by fine-tuning a second model
on DME scans.
All fluid types, IRF, SRF, and PED (fibrovascular, serous, and drusenoid),
within a total of 15 different semantic labels were segmented as part of a system
from DeepMind for OCT diagnosis and referral [27]. Therefore, a 3D U-net [28]
was developed that used nine contiguous slices as a context to segment the middle
slice. In addition to identify ambiguous regions, an ensemble of five instances of
the network was constructed by training with a different order over the inputs and
different random weight initializations. Networks were trained on 877 (Topcon)
and 152 (Spectralis) OCTs having sparse annotations, that is, three to five B-scans
per volume were manually annotated, an equivalent of approximately 20 fully an-
notated OCT volumes. However, the segmentation performance metrics have not
been reported.