94
S. Zorgui et al.
– BatchNorm reduces the covariance shift and allows each network layer to
learn a little independently of the others.
– Label Smoothing is a regularizing component applied to the loss formula to
prevent overfitting.
The InceptionV3 network is 42 layers deep. Therefore, the computational cost is
just around 2.5 higher than GoogLeNet’s [21]. In addition, the inception modules
are a novel and popular concept due to their smaller convolutions, which explains
the reduction in the number of parameters. The InceptionV3 model gathers more
information without impacting the computational speed thanks to its depth and
the various kernel sizes used in the convolution operations.
3.4 Transfer Learning
As shown in Fig. 3, transfer learning [22] is proposed in order to ensure better
performance of the model. The model needs lots of labeled images to be capable
of solving complex problems. This has proved to be challenging especially when
the available dataset is small. Transfer learning is a deep learning method, in
which a model developed for a task is reused for a second task. This technique
uses pre-trained models as a starting point for other medical imaging tasks given
the vast computational and time resources required to develop CNN models on
these problems.
3.5 Prediction Model
In the prediction phase, the RCM images test set are resized and provided to the
trained CNN. Our system calculates a prediction score for each test image after
resizing it and compares it with the threshold T equal to 0.5. The threshold value
is chosen that way due to the fact that we are performing a binary classification.
The classification condition is as follows: if the predicted score (PS) value of
the image test is lower than T then this RCM image is classified as lentigo and
conversely.
4 Experimental Validation
This section evaluates the validation of the proposed lentigo detection method on
real RCM data. In our work, the dataset is provided from Lab. Pierre Fabre. In
this experiment, the data include 428 RCM images which high spatial resolutions
and annotation on each image into two healthy and lentigo classes. The images
were acquired with a Vivascope 1500 apparatus. Each RCM image shows a field
of view of 500 × 500 µm with 1000 × 1000 pixels. A selection of 45 women aged
60 years were recruited. All participants have offered their informed consent to
the RCM skin test. We split these data into three main sets:
– A 314 images training set divided into two classes of 160 healthy images and
154 lentigo images.
S. Zorgui et al.
– BatchNorm reduces the covariance shift and allows each network layer to
learn a little independently of the others.
– Label Smoothing is a regularizing component applied to the loss formula to
prevent overfitting.
The InceptionV3 network is 42 layers deep. Therefore, the computational cost is
just around 2.5 higher than GoogLeNet’s [21]. In addition, the inception modules
are a novel and popular concept due to their smaller convolutions, which explains
the reduction in the number of parameters. The InceptionV3 model gathers more
information without impacting the computational speed thanks to its depth and
the various kernel sizes used in the convolution operations.
3.4 Transfer Learning
As shown in Fig. 3, transfer learning [22] is proposed in order to ensure better
performance of the model. The model needs lots of labeled images to be capable
of solving complex problems. This has proved to be challenging especially when
the available dataset is small. Transfer learning is a deep learning method, in
which a model developed for a task is reused for a second task. This technique
uses pre-trained models as a starting point for other medical imaging tasks given
the vast computational and time resources required to develop CNN models on
these problems.
3.5 Prediction Model
In the prediction phase, the RCM images test set are resized and provided to the
trained CNN. Our system calculates a prediction score for each test image after
resizing it and compares it with the threshold T equal to 0.5. The threshold value
is chosen that way due to the fact that we are performing a binary classification.
The classification condition is as follows: if the predicted score (PS) value of
the image test is lower than T then this RCM image is classified as lentigo and
conversely.
4 Experimental Validation
This section evaluates the validation of the proposed lentigo detection method on
real RCM data. In our work, the dataset is provided from Lab. Pierre Fabre. In
this experiment, the data include 428 RCM images which high spatial resolutions
and annotation on each image into two healthy and lentigo classes. The images
were acquired with a Vivascope 1500 apparatus. Each RCM image shows a field
of view of 500 × 500 µm with 1000 × 1000 pixels. A selection of 45 women aged
60 years were recruited. All participants have offered their informed consent to
the RCM skin test. We split these data into three main sets:
– A 314 images training set divided into two classes of 160 healthy images and
154 lentigo images.