Research Background Bladder cancer is the ninth most common malignancy worldwide,
with an estimated 430,000 new cases diagnosed annually. The standard diagnosis and
monitoring of bladder cancer rely on white light cystoscopy (WLC), with over 2
million cystoscopies performed annually in the United States and Europe. Due to the
high recurrence rate of bladder cancer, frequent monitoring and intervention are
necessary.
Early detection and complete resection of non-muscle invasive bladder cancer can
reduce recurrence and progression. However, up to 40% of patients with multifocal
disease do not achieve complete resection during the initial transurethral resection
of bladder tumor (TURBT). Many papillary tumors and flat lesions are difficult to
identify through WLC. There is an urgent need for cost-effective, non-invasive, and
user-friendly adjunct imaging technologies to address the diagnostic deficiencies of
WLC.
Recent advancements in deep learning-based automated image processing may provide
new solutions to the limitations of cystoscopy. Convolutional neural networks (CNNs)
possess the ability to learn complex relationships and integrate existing knowledge
into models, showing potential applications across various fields, including bladder
tumor diagnosis. We employed the HRNet algorithm, a convolutional neural network,
for enhanced bladder tumor detection.
Research Process 4.1 Patient Cohort Inclusion criteria: 1. The patients who had
bladder tumor and received WLC or TURBT, and the full-length surgery video is
available.
Exclusion criteria: 1. The video is too blurry to distinguish the normal bladder
wall and bladder tumor. 2. Lack of the appearance of bladder tumor before resection.
- Lack of informed consent.
Patient information in the videos will not be shown. Videos from the initially
recruited 200 bladder tumor patients will be used for algorithm development. Videos
from an additional 100 patients are used for algorithm validation.
4.2 Data Preprocessing To reduce the data volume, we extract the frame at a ratio of
1:4. Two urologists outline the boundary of bladder tumors in each frame seperately
and check for each other. AI algorithm is used to contour the same bladder tumors.
The outlines of bladder tumor annotated by urologists and algorithm are compared,
the IOU value, precision, sensitivity and false negative rate are analyzed.
4.3 Algorithm Development This study uses semantic segmentation to identify bladder
tumors in WLC. The D-LinkNet network structure used a pre-trained ResNet34 on the
ImageNet dataset as its encoder, with the central part utilizing dilated
convolutions with different dilation rates in a cascaded manner, and upsampling
performed using deconvolution. The original resolution of all images are 1920×1080,
downsampled by 2 to 960×540, and zero-padded in the width direction to obtain the
image with a resolution of 960×544. The RGB images of this size were normalized,
mean-subtracted, and variance-divided before being input into the network. The
images undergo five encoding processes, dilated convolutions, and five decoding
processes, ultimately producing a prediction result of 960×544, which was further
post-processed. In this research, the parameter settings were as follows: batch size
of 8, Adam optimizer, a learning rate of 0.001. The training environment was an
NVIDIA TITAN Xp GPU.
4.4 Results Interpretation The intersection over union (IOU) is a crucial standard
for evaluating single-frame image recognition capability in image recognition. When
IOU is above the threshold, it suggests that the model detects the object
successfully, indicating a true positive. When IOU can not reach the threshold, it
suggests that the model fails to detect the object, and indicating a false negative.
If a prediction appears without ground truth in the image, it is considered a false
positive. We calculate the model's sensitivity and precision in the test set. The
Dice coefficient measures the similarity between two samples. A higher average Dice
coefficient indicates a better detection performance of the model.
4.5 Observation Indicators
① Video annotation and classification: annotation status and RLN recognition
discernibility classification results for training and test set surgery videos; ②
After grouping and training, observe the model's sensitivity, precision, false
negative rate, false positive rate, and average Dice coefficient at IoU thresholds
of 0.1 and 0.5 in the test set under different discernibility groups.
4.6 Statistical Methods Analysis will be analyzed by R 4.0.2 software, the data will
be expressed as absolute numbers or percentages.
