A Comparison of Image Analysis Tools for Segmentation on SEM Micrographs – Zeiss ZEN Intellesis vs. Thermofisher AVIZO

Austenitic stainless steels are among the most widely used steels and overall metal materials in industry. Their specific mechanical properties, such as high corrosion resistance, formability, weldability and high strength along wide spread of temperatures, grant their irreplaceable utilization across different sectors from food and pharmaceutical industries to energy industries [1,2]. In view of the fact that in many applications, human lives are dependent on their right functionality, we need to precisely know their behavior under extreme conditions, which includes examination and evaluation of their microstructure and revealing any microstructural changes on time. Evaluating the microstructure of advanced steels is time and cost-consuming and requires skilled professional – metallographer. Therefore, many companies are trying to utilize artificial intelligence to enhance and accelerate this process and also to get rid of human errors. Different products adopt different approaches from user-friendly pre-trained neural networks ready to use to learning your own neural network with adjustable parameters. One of the most famous software on the field of artificial intelligence in metallography are Zen Intellesis by ZEISS [3] and Amira-Avizo 2D by Thermo Fisher Scientific [4]. How would these software stand against each other and is it possible to achieve similar results with pretrained neural networks generally used for image classification? In this work, we will compare performance of two commercial software used for image analysis, specifically segmentation of phases in micrographs of the austenitic stainless steel captured by scanning electron microscope (SEM). Moreover, we will try to achieve better results in terms of segmentation accuracy by fine-tuning pretrained neural networks used for image segmentation tasks. The comparison will be held on a dataset of SEM micrographs of austenitic stainless steel AISI 301LN, which is commonly used in automotive and railway industry. In this particular case, the specimen underwent low cycle fatigue test which induced martensitic transformation. Presence of the α´-martensite phase in the austenite matrix of the austenitic stainless steel is unfavorable as it devalues mechanical properties of the steel such as strength and ductility. Therefore, the steel is more likely to suffer from embrittlement. On top of that, assessment of the occurrence of the α´-martensite in the microstructure is challenge on its own as there is not many ways, how to separate α´-martensite phase and austenite phase in an optical light microscope (LOM) or a scanning electron microscope (SEM) as shown in the fig.1 [5,6]. One of the commonly used ways how to determine phases in steels is the electron backscatter diffraction (EBSD) analysis in SEM. EBSD phase maps are precise way how to determine phases but also come with significant drawbacks. One of the biggest drawbacks of EBSD in general is time-consumption, need for special detector in SEM and high demand on quality of the specimen surface. For these reasons, new method based on channeling contrast was developed, providing strong halo-like contrast around α´-martensite grains as shown in fig.2. When the phase contrast was obtained, a dataset of micrographs was created for the training purposes of neural networks [7,8]. As of machine learning part, target problem is to segment phases in the microstructure of austenitic stainless steel with a high precision and low time requirements. The image processing challenge with such dataset is the fact, that the halo is not corresponding with the true shape of α´-martensite grains as it reaches into the matrix. Therefore, real size of the α´-martensite grains is smaller than the surrounding halo. For these reasons, complex machine learning approaches with fine-tuned feature selections are required [8]. The problem with the evaluation of the results from different software is the fact, that both Amira-Avizo and Zen Intellesis SW offer different approaches to the image analysis. To meet the objectivity of this work, precision, time of learning phase as well as time of preprocessing of data will be taken into account. To further minimize the differences, parameters such as loss function or patch size will be set equally. Preliminary results have been already achieved and are displayed below in tab.1,2 and 3. The preliminary test was focused on the comparison of all models included in the software and how well they can cope with the dataset. In terms of pretrained neural networks, different architectures were chosen, to assess their suitability. In case of Amira-Avizo their base model is based on U-Net architecture with variable backbones, for our study, ResNet, VGG16 and VGG19 were chosen and further trained in 150 epochs. In contrast, base model of Zen Intellesis is based on VGG19 without option of choosing a backbone to make this software as user-friendly as possible. However, there is a possibility to choose number of feature selections. For a deeper insight, all options were trained to examine differences. For the metric purpose, the IoU score was adopted [9]. In conclusion, our initial results show that most of the tested solutions prove themselves with promising accuracy. Overall best result achieved pretrained neural network based on U-Net architecture with the IoU score of 0,544. Models of Amira-Avizo performed comparatively well with the ResNet backboned model achieving 0,462. For the Zen Intellesis, the results suggest that richer feature selection models need to be chosen for this task, as the microstructure is too demanding for the less complex models to perform well. We can expect that with further fine-tuning of selected model, the outcomes will be even better. In a further experiment, the evaluation of time will be taken into consideration as well as more software from different companies will be tested to create general overview of the performance of image analysis software [10]. SEM micrograph of a same area obtained by the circular backscatter electrons detector (CBS) situated bellow the pole piece (left) and conventional Everhart-Thornley secondary electrons detector (right). EBSD phase map (left) and corresponding BSE micrograph obtained under specific conditions demonstrating the strong contrast around α´-martensite phase (right). Comparison of IoU score of different image analysis software. An example of a set from Zen Intellesis with CBS micrograph (top), phase mask (lower left) and predicted phase map (lower right)