Annotation of the results obtained in 2020
1. New architectures of the segmentation and classification subsystems have been developed, which make it possible to reduce the number of false-positive cases and increase the accuracy of differential diagnostics through the implementation of more complex compositions of Siamese neural networks. As a new segmentation architecture, it is proposed to use a combination of 3D detection and segmentation inside bounding boxes, which can significantly improve the segmentation accuracy. As a new classification architecture, it is proposed to use a system consisting of three parallel classification channels with the next calculation of accuracy measures of the entire algorithm. The second channel uses an ensemble of 80 triplet neural networks such that every network can be viewed as a generalization of the Siamese network. A special ensemble training procedure has been developed, which allows us to construct triplets under condition of a significant imbalance in the training data. 2. A quite new approach to the interpretation and explanation of diagnostic results based with natural language is proposed. The main ideas behind the approach are: 1) the natural language explanation is presented in the form of a hierarchy of primitives and phrases that describe the form, structure, inclusions, contours and other peculiarities of nodules; 2) construction and training of simple classifiers that classify suspicious objects or their low-dimensional representation into classes corresponding to primitives such that each classifier corresponds to a peculiarity of the nodule; 3) implementation of the whole explanation algorithm in the form of two parts: the first part is an explanation model (LIME or SHAP) for selecting meaningful features from the object or its representation, and the second part, the key one, is a set of classifiers whose purpose is to combine the selected meaningful features with sentences in natural language. 3. Two new modifications of the adaptive deep forest are proposed. The first one is the transfer learning which aims to solve the classification problem. A solution to the problem of unsupervised domain adaptation is proposed in the project when representation of the initial vectors at each level of the deep forest cascade with adaptive weights is used instead of transforming the feature space to achieve a minimum measure of the domain proximity. The second modification is based on the imprecise representation of weights of training examples using imprecise statistical models, for example, the imprecise Dirichlet model or the contamination model, which leads to many weight distributions as a part of the unit simplex and many random forests such that one forest is selected for testing, which gives the maximum training classification error. The imprecise model parameter allows to adaptively control the robustness of deep forest cascades. 4. New models have been developed for evaluating the heterogeneous treatment effect in small patient groups. The main idea is to use analogs of the local interpretation models and linearization of the “treatment group” model in a local area of the analyzed patient. The use of the distance metric between data, defined as the average distance over the entire forest between the leaves of decision trees, in which these examples fall, is proposed. The average distance is determined over the entire random forest. It is proposed to perform local linearization of the response function estimate based on several close observations on the basis of a similarity matrix constructed for a random forest. The LASSO model is used to mitigate the effect of data redundancy. 5. The concept of functioning and transformation of the radiology departments in oncological medical centers in the context of the use of intelligent systems for diagnosing diseases has been developed. It has been shown that the activity of a radiologist is transformed into a cyclic process, which implies, in addition to analyzing and interpreting images, monitoring the verification of pathology, assigning class labels for machine learning, labeling the pathology and forming a database of the studied pathology. Procedures have been developed for introducing the practice of constantly updating the database with each new detected case when a doctor receives a prediction produced by the intelligent system and compares the result with his own interpretation of the pathology. Most of the results are published in various journals, including journals indexed by Scopus and Web of Science.

 

Publications

---

Annotation of the results obtained in 2018
A quite new approach was proposed for controlling deep forests as well as random forests, which allows us to achieve two main goals: 1) to develop a mechanism for the forest control in terms of its orientation to a solved machine learning problem, which will bring deep forests closer to the universality and flexibility of deep neural networks; 2) to improve the classification or regression accuracy of deep forests by introducing additional elements of control over the process of classification or regression. The main idea of the approach proposed in the project is to build such a weighting function defined on the set of decision trees in each random forest in order to minimize the classification error or, more generally, to minimize some predefined loss function. In this case, outputs of decision trees are the class probability distribution. Weights are trained using the same training set. In contrast to the original deep forest, which solves only the standard classification problem, the proposed approach allows solving various problems by defining a required loss function for outputs of decision trees. This brings deep forests closer to the generality and flexibility of deep neural networks without leading to the overfitting problem under a small amount of training data. The first function of weights is the linear weighted averaging of the probability distributions of classes at the output of decision trees. A final optimization problem for computing optimal weights is quadratic with linear constraints. To effectively solve the problem, a modification of the Frank-Wolfe algorithm is proposed, which takes into account that the weights are restricted by a single simplex. The proposed approach is implemented also for random survival forests which can be viewed as regression models. A weighted random survival forest is for the first time developed as a modification of the standard survival forest. According to the modification, the averaging of hazard functions at the output of each decision tree, which is used to calculate the random forest hazard function, is replaced by a weighted sum of these functions. The calculation of the optimal weights is also reduced to solving a quadratic optimization problem with linear constraints, which maximizes the concordance index or the C-index. For the well-known Primary Biliary Cirrhosis (PBC) Dataset, the C-index increment was almost 9%. The project proposed for the first time new classification models based on deep forests with a reduced number of weights based on subsets of “closely” located class probability distributions defined on the unit simplex. Weights are assigned not to trees and not examples, but to subsets of the probability distributions of classes obtained at the output of each decision tree for each example. As a model for partitioning the unit simplex, Walley's imprecise pari-mutuel model was chosen. According to this model, the unit simplex is divided into many subsets. Weights of the subsets of the probability distributions can be viewed as second-order probabilities over the subsets of the simplex. The idea of assigning the weights not to decision trees, not to examples of training data, but to subsets of the class probability distributions, which are simultaneously defined by the classification “ability” of trees and by how an example is typical for its class, is proposed for the first time. A new classification models based on the reducing the weight set has been developed in the project. Moreover, an attempt to obtain new effective models led to an unexpected new scientific result. The reduction of the weight set is the weight smoothing and restriction, which is the basis of the regularization and a reason for using the regularization terms in the objective function. The use of the proposed modification of the Frank-Wolfe algorithm made it possible to construct a variety of the deep forest models using various restrictions on the unit simplex of weights. Several well-known imprecise statistical models were used including the imprecise contamination model, the imprecise Dirichlet model, the imprecise pari-mutual model, Kolmogorov-Smirnov bounds, the constant odds ratio model. New robust metric distance models have been developed, which, on the one hand, use a general approach to the deep forest modification using trained weights, and, on the other hand, propose completely new ideas aimed at implementing a general approach to solve tasks within metric models on trees and random forests. The algorithms of Siamese and triple neural networks have been first implemented on the random and deep forest models. The first model is to compare the object pairs and to change the relative location of objects in the feature space taking into account their belonging to the same classes. To ensure the convexity of the loss function, the idea was first proposed to combine the Euclidean distance and the Manhattan distance in a single loss function. The second model is used when the class labels are unknown. It analyses semantically close and long distance objects. This is the first complete analog of Siamese neural networks implemented on random forests. The model proposes to control the proximity of objects using the weights of trees and to get a new training sample consisting of concatenated pairs of objects. The third model is an alternative to triple neural networks. The idea of its learning is quite similar. A new algorithm for contouring pathological lung formations based on the multiplanar reconstruction of computed tomography images in the DICOM format has been developed and implemented. An idea underlying the proposed algorithm is to consider planar images of computed tomography images as a single three-dimensional object. The task is reduced to finding the object boundaries in three-dimensional space. For segmentation of lungs, a method based on the threshold inclusion algorithm has been developed. A new approach has been applied to the segmentation of objects in the lung CT images, which takes into account various types of lung nodule location. A frame-by-frame fill algorithm, a root contour algorithm using dilatation, and density filtering for the lung CT images have been also developed and implemented.

 

Publications

---

Annotation of the results obtained in 2019
New modifications of the deep forest were proposed to implement the data discrimination, taking into account their classes in a feature space. The problem of implementing distance metric learning methods on the basis of the deep forests was solved. It was proposed to assign weights to decision trees in all random forests in order to reduce distances between pairs of examples from the same class and to increase distances between pairs of examples from different classes. A special contrastive loss function, including two different distance metrics, was introduced to obtain the standard quadratic optimization problem. Modifications of the deep forest are also proposed for implementing the transfer learning task, where a consensus measure based on the Shannon entropy and an average distance (mean discrepancy) between the source and target domains are used to calculate the tree weights. New models of deep survival forests have been developed to solve problems of survival analysis. Within the framework of the models, it was proposed to change the averaging procedure used to estimate a forest survival function based on the survival function at the output of decision trees, to use the concordance index (C-index) as an objective measure for constructing the optimization problem, and to replace the C-index with its approximate representation which is based on the well-known hinge-loss function application. A new approach was proposed for using sets of class probability distributions at the next level (layer) of the deep forest. Decision trees of the next level of the forest cascade are trained on the basis of an expanded training set, which is added by new generated class probability distributions from their sets. The increased number of training examples is compensated by updating the hyperparameters in accordance with a used imprecise statistical model, for example, the imprecise Dirichlet model or the Kolmogorov-Smirnov bounds. In order to deal with sets of class probability distributions at the output of decision trees, and to ensure robustness, a meta-model is proposed that determines optimal weights of decision trees. As part of the approach, new loss functions for classification and regression problems were introduced, which made it possible to reduce complex minimax optimization problems for calculating optimal tree weights to quadratic optimization problems. This approach is also implemented for the survival analysis problem, where confidence intervals were used for Nelson-Aalen estimates. A new approach is proposed that uses sets of distributions, which consists in assigning weights not to trees, but to subsets of class probability distributions in a special way, which makes the solution more flexible and reduces the number of weights as training parameters. A new random and deep forest architectures have been proposed, using many neural networks to improve the classification accuracy. In fact, neural networks perform non-linear transformation of class probability distributions in such a way as to ensure maximum classification accuracy at the output of a random forest or deep forest. Neural networks can be regarded as an extension of Siamese neural networks, since they implement the same functions. The idea of using such an architecture opens up new possibilities for implementing a wide variety of machine learning tasks, including the transfer learning, anomaly detection, etc. it was proposed to consider the task of differential diagnosis of oncological diseases, especially atypical cases of cancer, as a task of one-shot or few-shot learning because of a small amount of training examples for atypical cases. Siamese neural networks and Siamese deep forests were proposed to use as tools for implementing the one-shot or few-shot learning. To increase the accuracy of a diagnosis, a three-channel architecture of the classification system for lung nodules on the lung computed tomography scans was developed to make decision about a diagnosis of a patient. The first channel is a classifier based on deep forests. The other two channels are two Siamese neural networks (fully connected neural networks and convolutional neural networks). A new modification of the deep forest, called Adaptive Weighted Deep Forest, is proposed. In accordance with this modification, each example is assigned a weight at the next level of the forest cascade, depending on how correctly it was classified at this level. A larger weight is assigned to “bad” examples so that classifiers at the next levels try to classify it correctly. Two strategies for using weights are considered: weighted examples are chosen randomly for training the decision trees in accordance with their weight distribution, and weights are used in the implementation of the splitting procedure when training the decision trees. A three-channel lung segmentation system is proposed, where the first channel is implemented as a conventional image processing procedure, the second channel is a deep learning procedure using a 3D U-Net segmentation neural network, which is actually a duplication of the first channel, and the third channel is a deep learning procedure using a 2D U-Net segmentation neural network for special segmentation cases that are difficult to accomplish with the first two channels. This architecture achieves the main goal - to avoid cases of missed tumors. In addition, new models for evaluating the heterogeneous treatment effect were proposed to implement the concept of personalized medicine based on random and deep forests for cases where the number of elements in the treatment group is small. An efficient meta-algorithm, called Co-learner, was proposed to evaluate the conditional average treatment effect, which is based on the concatenation of feature vectors from the control and treatment groups and the generation of additional concatenated vectors. Most of the results are published in various journals, including journals indexed by Scopus and Web of Science. Certificates of registration of 2 software programs, 1 database and 2 patents for invention were obtained. The results of the project were widely covered in the press, where indicated that the project is being implemented with the support of the Russian Science Foundation, examples of articles on the project are: https://tass.ru/obschestvo/5995816 https://minobrnauki.gov.ru/ru/press-center/card/?id_4=901 https://www.popmech.ru/science/news-458242-uchyonye-nashli-novyy-sposob-diagnostiki-opuholey/#part0 https://lenta.ru/news/2019/01/26/20_seconds/?utm_source=yxnews&utm_medium=desktop https://www.technologynetworks.com/tn/news/ai-for-lung-cancer-diagnostics-314929 https://ecmiindmath.org/2019/03/20/an-intelligent-system-for-lung-cancer-diagnostics/

 

Publications

---