nikvo1 · nikvo1 · Jun 1, 2026 · May 31, 2026 · May 31, 2026 · May 31, 2026
diff --git a/docs/Biological tasks guidelines/Population growth curves.rst b/docs/Biological tasks guidelines/Population growth curves.rst
@@ -169,6 +169,8 @@ Step 8. Press **Calibrate**.
 
 This calibration algorithm is fast and should quickly provide the result optimal confidence threshold in the line below the **Calibrate** button.
 
+.. note:: If the accuracy isn't high enough, there are two ways of increasing it. You can finetune NuclePhaser model using `Colab notebook <https://colab.research.google.com/drive/1hKMVQqYS0I_GrkYvdz23tPc8FCv2oJvh?usp=sharing>`_ or use :doc:`TTA </General information/Test-time augmentations (TTA)>`.
+
 Step 2: Applying the calibrated model to a stack of images
 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 
@@ -215,8 +217,6 @@ Surprisingly, the most influential factor on waiting time is **number of objects
 It's due to the fact that postprocess algorithm (NMS or NMM) has O(n\ :sup:`2`) `notation <https://en.wikipedia.org/wiki/Big_O_notation>`_, so it can take much more time than applying the deep learning algorithm itself.
 In our practice, inference time for one image exceeds 10-20 minutes when there are close to 100,000 objects on an image.
 
-.. hint:: You can track the process in the command line that you used to initiate the Napari (not available in standalone application). It provides an estimated time to process the whole stack.
-
 .. figure:: ../Images/Stack_CLI.jpg
         :scale: 50 %
         :align: center

diff --git a/docs/General information/Confidence threshold calibration.rst b/docs/General information/Confidence threshold calibration.rst
@@ -97,3 +97,12 @@ For more information, see :doc:`Calibrate with DAPI widget page </Widgets/Calibr
 Two metrics are generated during tests: `MAPE <https://en.wikipedia.org/wiki/Mean_absolute_percentage_error>`_ and prediction-ground truth scatterplot.
 The smaller the MAPE, the better. The closer predictions to the red line on scatterplot, the better.
 For more detailed information about the metrics, see our `paper <https://www.biorxiv.org/content/10.1101/2025.05.13.653705v1>`_.
+
+What if accuracy isn't high enough?
++++++++++++++++++++++++++++++++++++
+
+If the accuracy of the model on your specific use case is considered low, there are certain options to increase it.
+You can test another model from NuclePhaser model zoo. In our practice, small models sometimes better than large ones!
+Also, you can try :doc:`TTA (test-time augmentations) </General information/Test-time augmentations (TTA)>`. It is a way of increasing accuracy by sacrificing inference time.
+Finally, you can finetune NuclePhaser model on you specific use case using `Colab notebook <https://colab.research.google.com/drive/1hKMVQqYS0I_GrkYvdz23tPc8FCv2oJvh?usp=sharing>`_.
+However, it requires fluorescent nuclear stain.
diff --git a/docs/General information/Test-time augmentations (TTA).rst b/docs/General information/Test-time augmentations (TTA).rst
@@ -0,0 +1,47 @@
+Test-time augmentations (TTA)
+=============================
+
+Make sure you are familiar with :doc:`Confidence threshold calibration </General information/Confidence threshold calibration>`.
+
+What are test-time augmentations (TTA)?
+Put simply, they are a way to potentially increase accuracy by sacrificing inference time.
+
+Imagine you have a calibration image, and the model has an error of 10% on it.
+But what if, instead of using this native image, it would be slightly changed: contrasted, magnified or something else?
+It is possible that model would be better at these augmented images.
+For example, cells on calibration image are small, and resizing the image 1.5x resulted in error of 8%.
+Or even better, when we average the results from native and magnified image, their respective errors cancel each other, and the result error can be even lower!
+
+This is exactly what TTA are about.
+On calibration stage, instead of running single calibration on native image, it gets augmented in the following ways:
+
+**resize_1.5x** - image is increased in size 1.5x
+
+**resize_2x** - image is increased in size 2x
+
+**clahe** - `CLAHE (contrast-limited adaptive histogram equalization) <https://en.wikipedia.org/wiki/Adaptive_histogram_equalization#Contrast_Limited_AHE>`_ is applied to the image.
+
+**gamma 1.5** - `gamma contrast <https://en.wikipedia.org/wiki/Gamma_correction>`_ with coefficient 1.5 is applied to the image
+
+**invert** - all image intensities are inverted (255 becomes 0, 0 becomes 255 and so on)
+
+**median_3** - `median filter <https://en.wikipedia.org/wiki/Median_filter>`_ with kernel size 3 is applied to the image
+
+**bilateral_10** - `bilateral filter <https://en.wikipedia.org/wiki/Bilateral_filter>`_ with sigma value 10 is applied to the image
+
+**sharpen** - `unsharp mask <https://en.wikipedia.org/wiki/Unsharp_masking>`_ is applied to the image.
+
+Best threshold is found for each augmentation independently.
+The image is used for calibration with each augmentation applied, and then it gets tested on each possible combination of these augmentations.
+
+.. note:: How combination of augmentations work? Each augmentation is applied individually and results from them are averaged, not all of augmentations are applied simultaneously.
+
+The main tradeoff is inference time. For example, the combination of 5 augmentations decreases the error from 10% to 9%.
+It means that for increasing accuracy by 1% you will have to spend 5x more time on inference.
+
+You can calibrate with TTA using :doc:`Calibrate with points widget </Widgets/Calibrate with points>`.
+The widget creates a "TTA" subfolder in the given folder for storing the calibration results.
+metadata_TTA.txt file is stored in that subfolder.
+All inference (predict on single image, predict on 1-stack, predict on 2-stack) widgets have "Use TTA" checkbox and a field for selecting .txt file.
+Select your metadata_TTA.txt calibration file, and widget will perform TTA automatically.
+Widget will create a points layer for each augmentation, and the result .csv or .xlsx file will contain averaged counting results.
diff --git a/docs/Installation/Installation.rst b/docs/Installation/Installation.rst
@@ -1,7 +1,15 @@
 Installation instructions
 =========================
 
-Option 1: Using Anaconda (recommended)
+Option 1 (simple): Using standalone napari app
++++++++++++++++++++++++++++++++++++++++++++++++
+
+.. note:: The only disadvantage of this option is inability to install GPU-accelerated version of torch
+1. Download and install napari as standalone app using `installation instructions <https://napari.org/stable/getting_started/installation.html#installation-bundle-conda>`_
+
+2. Search, download and install napari-nuclephaser plugin by opening the app, navigating to Plugins window and choosing Install/Uninstall plugins.
+
+Option 2 (recommended): Using Anaconda
 ++++++++++++++++++++++++++++++++++++++
 
 We recommend installation using `Anaconda Distribution <https://www.anaconda.com/>`_
@@ -14,7 +22,7 @@ We recommend installation using `Anaconda Distribution <https://www.anaconda.com
 
 .. code-block:: python
 
-   conda create --name napari-env anaconda
+   conda create --name napari-env python=3.10
 
 4. Activate new environment using command
 
@@ -48,7 +56,7 @@ We recommend installation using `Anaconda Distribution <https://www.anaconda.com
 
 Initialize plugin's widgets by opening Plugins window and choosing NuclePhaser.
 
-Option 1 advanced: installation with GPU
+Option 2 advanced: installation with GPU
 ++++++++++++++++++++++++++++++++++++++++
 
 If you have `NVIDIA GPU with CUDA <https://developer.nvidia.com/cuda-gpus>`_, you can significantly increase plugin's speed.
@@ -75,12 +83,3 @@ To install GPU-powered version of the plugin, you first need to do all the steps
 .. code-block:: python
 
    pip3 install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
-
-Option 2: Using standalone napari app (simpler)
-+++++++++++++++++++++++++++++++++++++++++++++++
-
-.. note:: Downsides of this option are: you can't install GPU-powered version and you will not have CLI (Command Line Interface) that prints detailed progress. Otherwise, it's the same.
-
-1. Download and install napari as standalone app using `installation instructions <https://napari.org/dev/tutorials/fundamentals/installation_bundle_conda.html>`_
-
-2. Search, download and install napari-nuclephaser plugin by opening the app, navigating to Plugins window and choosing Install/Uninstall plugins.
diff --git a/docs/Widgets/Calibrate with points.rst b/docs/Widgets/Calibrate with points.rst
@@ -53,6 +53,9 @@ For example, if you have an image 6400x6400 pixels, and Division size = 640, you
 **Calibration proportion** determines which part of small images array will be used for calibration, and which part - for test.
 If you have an array of 100 small images and Calibration proportion = 0,1, 10 of those images will be used for calibration, 90 - for test.
 
+**Test with TTA** checkbox is used for running calibration and test with TTA (test-time augmentations).
+Learn more at :doc:`page about TTA </General information/Test-time augmentations (TTA)>`.
+
 **Save folder** is used for selecting a folder in which the calibration plot and metadata.txt files will be saved.
 Inside this folder, a subfolder will be created with Expereiment name.
 

diff --git a/docs/Widgets/Count points on single image.rst b/docs/Widgets/Count points on single image.rst
@@ -1,5 +1,5 @@
-Count points on single image
-============================
+Count points on images
+======================
 
 Description
 +++++++++++

diff --git a/docs/Widgets/Predict on 1-stack.rst b/docs/Widgets/Predict on 1-stack.rst
@@ -35,6 +35,12 @@ Larger models can be downloaded on `NuclePhaser GitHub page <https://github.com/
 Confidence threshold is the **most important paramter** for the task of counting objects.
 Learn more about how to find the optimal threshold for your specific use case at :doc:`Confidence threshold calibration page </General information/Confidence threshold calibration>`.
 
+**Use TTA** checkbox is used for running inference with TTA (test-time augmentations). Requires passing metadata_TTA.txt created by Calibrate with points widget.
+Learn more at :doc:`page about TTA </General information/Test-time augmentations (TTA)>`.
+
+**TTA metadata file** field is used for passing the metadata_TTA.txt file created by Calibrate with points widget.
+Learn more at :doc:`page about TTA </General information/Test-time augmentations (TTA)>`.
+
 **Save result** checkbox is used for selecting whether you need counting results or not.
 If the task is to count the number of objects on each frame, check this box.
 It will create a subfolder with **Experiment name** at **Save folder** location with .csv and/or .xlsx file with counting results, as well as metadata.txt file with all the parameters for exact reproduction of results.

diff --git a/docs/Widgets/Predict on 2-stack.rst b/docs/Widgets/Predict on 2-stack.rst
@@ -35,6 +35,12 @@ Larger models can be downloaded on `NuclePhaser GitHub page <https://github.com/
 Confidence threshold is the **most important paramter** for the task of counting objects.
 Learn more about how to find the optimal threshold for your specific use case at :doc:`Confidence threshold calibration page </General information/Confidence threshold calibration>`.
 
+**Use TTA** checkbox is used for running inference with TTA (test-time augmentations). Requires passing metadata_TTA.txt created by Calibrate with points widget.
+Learn more at :doc:`page about TTA </General information/Test-time augmentations (TTA)>`.
+
+**TTA metadata file** field is used for passing the metadata_TTA.txt file created by Calibrate with points widget.
+Learn more at :doc:`page about TTA </General information/Test-time augmentations (TTA)>`.
+
 **Save result** checkbox is used for selecting whether you need counting results or not.
 If the task is to count the number of objects on each frame, check this box.
 It will create a subfolder with **Experiment name** at **Save folder** location with .csv and/or .xlsx file with counting results, as well as metadata.txt file with all the parameters for exact reproduction of results.

diff --git a/docs/Widgets/Predict on single image.rst b/docs/Widgets/Predict on single image.rst
@@ -52,6 +52,12 @@ Select it if you want to generate Napari Shapes layer with a rectangle for each
 **Show confidence** checkbox is used to add confidence scores for each bounding box.
 Works only if **Generate bbox** checkbox is active.
 
+**Use TTA** checkbox is used for running inference with TTA (test-time augmentations). Requires passing metadata_TTA.txt created by Calibrate with points widget.
+Learn more at :doc:`page about TTA </General information/Test-time augmentations (TTA)>`.
+
+**TTA metadata file** field is used for passing the metadata_TTA.txt file created by Calibrate with points widget.
+Learn more at :doc:`page about TTA </General information/Test-time augmentations (TTA)>`.
+
 **Confidence threshold** field is used to set up a confidence threshold for the YOLO model.
 Confidence threshold is the **most important paramter** for the task of counting objects.
 Learn more about how to find the optimal threshold for your specific use case at :doc:`Confidence threshold calibration page </General information/Confidence threshold calibration>`.

diff --git a/docs/conf.py b/docs/conf.py
@@ -9,7 +9,7 @@
 project = "napari-nuclephaser"
 project_copyright = "2025, Nikita Voloshin"
 author = "Nikita Voloshin"
-release = "0.0.5"
+release = "0.3.0"
 
 # -- General configuration ---------------------------------------------------
 # https://www.sphinx-doc.org/en/master/usage/configuration.html#general-configuration

diff --git a/docs/index.rst b/docs/index.rst
@@ -34,6 +34,7 @@ To learn more about the algorithms behind the project, navigate to general infor
    General information/Object detection overview.rst
    General information/Sliced inference overview.rst
    General information/Confidence threshold calibration.rst
+   General information/Test-time augmentations (TTA).rst
 
 To learn more about widgets and their functions, navigate to the widget page you need:
 

diff --git a/pyproject.toml b/pyproject.toml
@@ -30,7 +30,7 @@ dependencies = ["setuptools==80.10.2",
 		"ultralytics==8.3.141",
 		"yolov5",
 		"magicgui",
-		"sahi==0.11.36",
+		"sahi==0.12.0",
 		"scikit-image",
 		"torch==2.8.0",
 		"pathlib",

diff --git a/src/napari_nuclephaser/calibrate_dapi.py b/src/napari_nuclephaser/calibrate_dapi.py
@@ -1,5 +1,6 @@
 import os
 import pathlib
+import warnings
 from datetime import datetime
 
 import cv2
@@ -18,6 +19,9 @@
 
 from napari_nuclephaser.utils import create_unique_subfolder, initialize_model
 
+warnings.filterwarnings(action="ignore", category=FutureWarning)
+warnings.filterwarnings(action="ignore", category=UserWarning)
+
 matplotlib.use("Agg")
 # cuda device check
 cuda_available = "cuda:0" if cuda.is_available() else "cpu"
@@ -209,6 +213,7 @@ def split_image(image, size):
             postprocess_match_metric=Match_metric,
             postprocess_match_threshold=Intersection_threshold,
             verbose=0,
+            force_postprocess_type=True,
         )
 
         dapi_count = int(len(dapi_result.object_prediction_list))
@@ -224,6 +229,7 @@ def split_image(image, size):
             postprocess_match_metric=Match_metric,
             postprocess_match_threshold=Intersection_threshold,
             verbose=0,
+            force_postprocess_type=True,
         )
 
         detection_confidences = []
@@ -287,6 +293,7 @@ def split_image(image, size):
             postprocess_match_metric=Match_metric,
             postprocess_match_threshold=Intersection_threshold,
             verbose=0,
+            force_postprocess_type=True,
         )
 
         dapi_count = len(dapi_result.object_prediction_list)
@@ -303,6 +310,7 @@ def split_image(image, size):
             postprocess_match_metric=Match_metric,
             postprocess_match_threshold=Intersection_threshold,
             verbose=0,
+            force_postprocess_type=True,
         )
 
         phase_count = len(phase_result.object_prediction_list)

diff --git a/src/napari_nuclephaser/calibrate_known_number.py b/src/napari_nuclephaser/calibrate_known_number.py
@@ -101,6 +101,7 @@ def calibrate_with_known_number(
         postprocess_type=Postprocess,
         postprocess_match_metric=Match_metric,
         postprocess_match_threshold=Intersection_threshold,
+        force_postprocess_type=True,
     )
     result = result.to_coco_predictions()
     print("Prediction is complete!")