Performance Review of Retraining and Transfer Learning of DeLTA2 for

Image Segmentation for Pseudomonas Fluorescens SBW25

Beate Gericke

1 a

, Finn Degner

, Tom H

uttmann

, S

oren Werth

3 b

and Carsten Fortmann-Grote

1 c

Max Planck Institute for Evolutionary Biology, Pl

on, Germany

Technische Hochschule L

ubeck, L

ubeck, Germany

Berliner Hochschule f

ur Technik, Berlin, Germany

ﬁ

Keywords:

Deep Neural Networks, Image Analysis, Supervised Learning, Cell Size, Jaccard Index, Intersection Over

Union, Balanced Accuracy.

Abstract:

High throughput microscopy imaging yields vast amount of image data, e.g. in microbiology, cell biology,

and medical diagnostics calling for automated analysis methods. Despite recent progress in employing deep

neural networks to image segmentation in a supervised learning setting, these models often do not meet the

performance requirement when used without model reﬁnement in particular when cells accumulate and overlap

in the image plane. Here, we analyse segmentation performance gains obtained through retraining and through

transfer learning using a curated dataset of phase contrast microscopy images taken of individual cells and cell

accumulations of Pseudomonas ﬂuorescens SBW25. Both methods yield signiﬁcant improvement over the

baseline model DeLTA2 (O’Conner et al. PLOS Comp. Biol 18, e1009797 (2022)) in intersection–over–union

and balanced accuracy test metrics. We demonstrate that (computationally cheaper) transfer learning of only

25% of neural network layers yields the same improvement over the baseline as a complete retraining run.

Furthermore, we achieve highest performance boosts when the training data contains only well separated cells

even though the test split may contain cell accumulations. This opens up the possibility for a semi–automated

segmentation workﬂow combining feature extraction techniques for ground truth mask generation from low

complexity images and supervised learning for the more complex data.

1 INTRODUCTION

Advances in technology allow modern biology to

gather increasing amounts of data (Stephens et al.,

2015). Since manual analysis cannot keep up with

the increase of gathered data, automated data analysis

methods become more and more important. Besides

genomic and proteomic sequence data, microscopy

imaging is among the major data sources (Peng, 2008)

in various areas such as microbiology, cell biology or

neurobiology.

Here we are interested in applications in the do-

main of evolutionary microbiology. Pseudomonas

ﬂuorescens is an important model organism in this

ﬁeld. Interest in our strain SBW25 (Rainey and Trav-

isano, 1998; Silby et al., 2009) originates from its

https://orcid.org/0009-0003-5777-7945

https://orcid.org/0009-0001-7936-2391

https://orcid.org/0000-0002-2579-5546

potential beneﬁts for host plants, e.g. in an agricul-

tural context (Thompson et al., 1993). Evolution-

ary research on SBW25 aims at characterizing genet-

ically modiﬁed derivatives in terms of cell size and

shape, metabolism, growth dynamics, evolutionary

ﬁtness, and ecology. Complementing genotyping via

targeted and whole genome sequencing, timelapse–

microscopy of growing cell colonies provides cell

phenotype data such as cell size and shape, as well

as growth rates as a function of time. To this end, mi-

croscopy images are segmented to identify individual

cells followed by cell counting, size and shape analy-

sis, classiﬁcation and tracking.

Convolutional Neural Networks (CNNs) have

been demonstrated to be a viable solution for com-

puter vision tasks (LeCun et al., 2015), such as

segmentation, object recognition and classiﬁcation.

Here, our focus is on image segmentation, i.e. the

task to identify if a pixel belongs to a cell or not. Seg-

mentation is usually the ﬁrst task in a image analysis

Gericke, B., Degner, F., Hüttmann, T., Werth, S. and Fortmann-Grote, C.

Performance Review of Retraining and Transfer Learning of DeLTA2 for Image Segmentation for Pseudomonas Fluorescens SBW25.

DOI: 10.5220/0012316300003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 273-280

ISBN: 978-989-758-688-0; ISSN: 2184-4305

273

workﬂow. Errors in segmentation will propagate and

potentially amplify in later workﬂow stages, such as

tracking or cell counting.

For biological and medical image segmenta-

tion, various implementations have been devel-

oped, such as Cellpose (Stringer et al., 2021),

DeLTA2 (O’Connor et al., 2022), MiSiC (Panigrahi

et al., 2021), and more recently Omnipose (Cutler

et al., 2022). Despite promising results in their re-

spective realm of application and with respect to their

training and test data, we found their accuracy suf-

fered when applied to image data from our model or-

ganism Pseudomonas ﬂuorescens SBW25.

Generic models achieving suboptimal results

when applied to special datasets is a known phe-

nomenon and was discussed e.g. in (Vaswani et al.,

2017; Campello et al., 2021; Brown et al., 2020).

The main reason for the failure of generic models is

the under–representation of the special dataset in the

training data of the model (Ma et al., 2023). This

in turn is mainly caused by the data generation and

training process being inherently difﬁcult and costly

in time and human resources, if automated ground

truth labeling is not available. This process often im-

plies manual editing of individual images to create an

accurate ground truth. In the case of bacteria, label-

ing of cell boundaries becomes ambiguous to a certain

degree if cells touch or overlap, increasing the human

work load.

Transfer Learning (TL) is a proven method to

adapt generic models to specialized datasets (Weiss

et al., 2016; Kim et al., 2022; Yu et al., 2022; Iman

et al., 2023). In the case of CNNs this can be un-

derstood intuitively realizing that a CNN effectively

acts as a series of ﬁlters. While early layers (ﬁlters)

are mostly sensitive to speciﬁc aspects of the image

(e.g. edges) later layers are sensitive to generic and

abstract features (Chollet, 2018). Hence, retraining

only the early layers transfers a model’s ability to seg-

ment images of one class of objects to another class

in the case where these classes (such as bacteria of

different species or strains) share certain generic fea-

tures (such as their overall shape), but differ in more

subtle features (e.g. length or curvature). In the limit

of retraining all layers, transfer learning and complete

retraining coincide. TL becomes especially useful if

not much data for a retraining is available or if com-

pute time for complete retraining is limited.

In the following, we describe how we performed

and evaluated complete retraining and TL on a cu-

rated dataset of microscopy images taken from Pseu-

domonas ﬂuorescens SBW25 cells. We quantify the

model’s ability to segment our images in terms of

performance metrics Balanced Accuracy (BA) and

Intersection–over–Union (IoU). As a baseline model

for comparison, we employ the trained segmentation

model from the DeLTA2 software (O’Connor et al.,

2022). DeLTA2 is also the base model for our TL

experiments. We demonstrate that both complete re-

training and TL yield a signiﬁcant improvement over

the baseline. Moreover, we ﬁnd that transfer learning

of only the top 25% of layers of the DeLTA2 model

gives the same quantitative improvement over the un-

modiﬁed DeLTA2 model as complete retraining, un-

derlining the effectiveness of this method.

Secondly, we found that transfer learning yields

accurate segmentation even for rather complex and

dense cell accumulations although the training data

contains only images with few and well separated

cells. This result is remarkable insofar as the label-

ing of such low complexity images can be done in

an automatized way using classical feature based seg-

mentation techniques (e.g. thresholding or edge de-

tection). This ﬁnding opens the possibility of largely

automated (unsupervised) segmentation, where man-

ual labeling of the training data becomes unnecessary.

2 DATASET CREATION

2.1 Microscopy

Our dataset consists of 34 time series of phase con-

trast microscopy images from growing Pseudomonas

ﬂuorescens cultures growing in different media. Data

was taken on a Zeiss Axio Imager Z2 with 100x mag-

niﬁcation. Each time series consists of 10-20 images

taken at ﬁxed time intervals. All series start shortly af-

ter the moment of inoculation with a single cell. The

series end at different timepoints, usually when one or

multiple communities of closely adjacent cells have

emerged. In total, 412 images of 2048x2048 pixels

have been included in this study. We refer to these

images as the MPB dataset. They are available from

(Fortmann-Grote and other, 2023).

2.2 Manual Masking

Manual creation of masks (i.e. black–white im-

ages with white denoting pixels that belong to a

cell and black pixels belonging to the background)

was performed using Adobe Photoshop version 23.5.1

(Adobe) and Afﬁnity Photo version 1.10.6 (SerifLtd)

using a touchscreen device and a graphics tablet (Wa-

com), respectively. Before masking, images were in-

dividually adjusted in brightness and contrast to en-

hance visibility of cell regions to the human eye. Ad-

justed images were disposed afterwards and not used

BIOIMAGING 2024 - 11th International Conference on Bioimaging

274

for training. We marked the outline of cells with a

tablet pen. Touching or overlapping cells were artiﬁ-

cially separated by a one pixel thick boundary. The

set of 412 images and 412 masks will be referred to

as complete dataset henceforth.

Additionally, we generated masks for a subset of

226 images representing the ﬁrst 5 time steps in our

series. In this early times of the growth experiment,

cell numbers are low and the few cells are isolated.

Ground truth labeling could hence be carried out by

intensity thresholding. We used the “modiﬁed Iso-

Data thresholding method” implemented in MicrobeJ

version 5.13n (8) – beta (Ducret et al., 2016) with off-

set threshold 180, stack histogram thresholding and

bicubic resampling (p = 0.5). The set of early time

images and masks is termed the partial dataset.

2.3 Training Validation Test Split

For training with the complete MPB dataset one time

series with 16 images was set aside, the remaining 33

time series were split into 7 time series (88 images)

for validation and 26 time series (308 images) for neu-

ral network training using backpropagation. For train-

ing with the partial MPB dataset 88 images were used

for training and 49 images for validation. These 49

validation images are taken from the complete dataset

to prevent leaky validation.

3 THE DeLTA2 BASELINE

MODEL

DeLTA2 is a deep learning segmentation and track-

ing pipeline for two–dimensional time-lapse mi-

croscopy (O’Connor et al., 2022). The DeLTA2 seg-

mentation model was trained on phase contrast im-

ages of Escherichia coli cells and achieves impressive

segmentation accuracy of the order 99%. Since E.

coli bacteria are similar to Pseudomonas ﬂuorescens

SBW25 cells in shape and size, we chose the DeLTA2

segmentation model as our baseline and as the start-

ing point for retraining and TL. DeLTA2 has a U–Net

architecture (Ronneberger et al., 2015) featuring two

symmetric legs: Input images are fed into the con-

traction leg with 5 sets of alternating convolution and

max–pooling layers. The expansion leg consists of 4

levels of alternating upsampling and convolution lay-

ers with additional concatenation of output from the

corresponding level on the contraction leg. In total,

the DeLTA2 model has 36 layers.

When evaluated on the complete MPB dataset,

we found two major issues with the DeLTA2 model.

(a) (b) (c)

Figure 1: Raw image data (a), manually created ground

truth binary mask (b), and (c) segmentation with DeLTA2

for a cell colony in an example image from our MPB train-

ing dataset. Colored areas in (c) indicate true positive

(white), false positive (green), false negative (red) and true

negative (black) pixel segmentation.

These are is illustrated in Fig. 1 showing a phase con-

trast image (1a) of Pseudomonas ﬂuorescens SBW25

zoomed in on a cell cluster of 48 closely neighbor-

ing cells, the ground truth mask (1b) and the DeLTA2

segmentation result (1c). The colored areas mark the

DeLTA2 segmentation. The ﬁrst issue is that long

cells are often split into two by the model, adding

a narrow intercellular space (red pixels, false posi-

tive predictions). Secondly, the model wrongly clas-

siﬁes pixels outsite the true cell perimeter as being

part of the cell (green pixels, false positives) in nu-

merous cases of smaller and isolated cells. In quan-

titative terms, DeLTA2 achieves only ≈ 90% BA and

≈ 65% IoU for our data (see Sec. 4.2), far below the

reported values for the original DeLTA2 test evalua-

tions. This ﬁnding is clearly unsatisfactory for our use

case, hence the motivation for retraining DeLTA2 on

our data.

4 RETRAINING EXPERIMENTS

We consider how the DeLTA2 model performs on

phase contrast images of Pseudomonas ﬂuorescens

SBW25 with and without retraining. The combina-

tion of our two curated training datasets (complete

and partial and the two considered training methods

(complete retraining and transfer learning with only a

subset of neural network layers) deﬁnes four experi-

ments:

1. Complete retraining with complete dataset

2. Transfer learning with complete dataset

3. Complete retraining with partial dataset

4. Transfer learning with partial dataset

Performance Review of Retraining and Transfer Learning of DeLTA2 for Image Segmentation for Pseudomonas Fluorescens SBW25

275

In all cases, the DeLTA2 segmentation model is the

starting point. In the case of complete retraining, all

network weights are initialized with random values,

while in the case of TL, the network was initialized

with the published pre–trained DeLTA2 model. After

training, all four cases are evaluated by calculating

BA and IoU on the respective test split.

4.1 Training Parameters

DeLTA2 uses weight maps to emphasize important

parts of an image. These were regenerated for all

training data using a utility function from the DeLTA2

repository. An example weight map is included in the

supplementary material (Fortmann-Grote and other,

2023). The complete retraining of the DeLTA2 neu-

ral network was performed over 600 epochs with 300

steps in each epoch. Early stopping with a patience

of 50 epochs was applied to mitigate overﬁtting. The

Adam optimizer was employed with with a learning

rate of 0.0001. Backpropagation used the pixel wise

weighted binary cross entropy (part of the DeLTA2

software) as loss function.

Transfer learning was run over 10 epochs with 300

steps and otherwise unchanged parameters. Training

started from DeLTA2 pre-trained model. Retraining

2, 9, 18, 27, or 36 layers results in 5 distinct TL mod-

els in total.

4.2 Metrics

In order to compare the performance of our differ-

ent models, we employ two performance metrics,

BA (Minaee et al., 2020) and IoU. These are deﬁned

in terms of the number of true positive, true negative,

false positive, and false negative pixel classiﬁcations

per image, TP, TN, FP, and FN,

TP =

A ∩ B

FP =

A ∩ ¬B

(1)

FN =

¬A ∩ B

TN =

¬A ∩ ¬B

, (2)

with A and B being the set of pixels labeled as belong-

ing to a cell in the model prediction and in the ground

truth, respectively.

Then, BA is deﬁned as

Balanced Accuracy =

Recall + Speciﬁcity

(3)

with:

Recall =

TP + FN

Speciﬁcity =

TN + FP

(4)

BA was suggested as a suitable metric for skewed

(imbalanced) datasets(Garc

ıa et al., 2009).

IoU is the ratio of the number of pixels correctly

predicted as “cell” (intersection between prediction

and ground truth) and the number of pixels being la-

beled as “cell” in either prediction or ground truth or

both (union of prediction and ground truth):

IoU =

| A ∩ B |

| A ∪ B |

TP + FN + FP

(5)

IoU is considered a very robust test metric (Minaee

et al., 2020) and can be shown to be the strictest eval-

uation metric for reasonably well performing classiﬁ-

cation tasks.

5 RESULTS

5.1 Retraining

The model resulting from a complete retraining with

the complete MPB dataset achieves a mean BA(IoU)

of 0.91(0.75). Compared to 0.85(0.65) achieved by

the DeLTA2 pretrained model, this is a signiﬁcant im-

provement. In Fig. 2 we show a color coded compar-

ison between model predicted masks from test data

and the ground truth. Mask pixels correctly identi-

ﬁed as belonging to a cell are highlighted in white

(true positives), false positives are colored in green,

false negatives are colored red and true negatives in

black. While the baseline model DeLTA2 (Fig. 2a)

yields a high false–positive rate (indicated by green

areas), retraining on either the complete (Fig. 2b) or

the partial (Fig. 2c) MPB dataset signiﬁcantly reduces

the false–positive rate. On the other hand, the false–

negative rate slightly increases compared to the base-

line (indicated by red areas) but a net improvement in

overall BA and IoU remains. Remarkably, both vari-

ants of completely retrained model (using the com-

plete dataset and using the partial dataset) yield sim-

ilar segmentations; the respective masks are indistin-

guishable by eye.

5.2 Transfer Learning

Table 1 tabulates the averaged balanced accuracy and

IoU values obtained from various TL models differing

in the numbers of retrained layers. The conﬁdence

intervals are taken as the twofold standard deviation

over three independently trained models and over all

images in the timeseries. All TL models improve the

segmentation performance compared to the baseline

model DeLTA2 (corresponding to 0 retrained layers).

Within the two sigma error margins all retrained mod-

els achieve the same performance gain. Looking at

the average values only, we ﬁnd a slight improvement

when retraining nine or more layers compared to the

two layer case, but the difference is not signiﬁcant.

BIOIMAGING 2024 - 11th International Conference on Bioimaging

276

(a) DeLTA2 model evaluated on MPB

test dataset

(b) MPB model completely retrained on

complete MPB dataset

partial MPB dataset

Figure 2: Color coded pixel map of true positive (white), false positive (green), false negative (red), and true negative (black)

segmentation masks for a selected image of the MPB test dataset.

Figure 3: Balanced accuracy (left) and IoU (right) vs. image timestamp for transfer learning with varying number of retrained

layers for the complete MPB timeseries dataset. The shaded areas denote the min/max range over three independently trained

models. Horizontal lines next to y-axis indicate total mean values over the timeseries according to Tab. 1.

Table 1: Mean balanced accuracy and mean IoU for transfer

learning models trained on the complete MPB dataset and

with varying number of retrained layers. 0 retrained layers

corresponds to the original DeLTA2 model. Error margins

are computed as 2 times the standard deviation over all test

images and over three independently trained model repli-

cates.

retrained layers mean BA mean IoU

0 0.85 0.65

2 0.88 ± 0.04 0.73 ±0.06

9 0.89 ± 0.04 0.76 ±0.06

18 0.91 ± 0.03 0.79 ±0.04

27 0.90 ± 0.02 0.78 ±0.04

36 0.90 ± 0.03 0.77 ±0.05

We ﬁnd the same pattern when we group the test im-

ages according to their timestamp in the microscopy

timeseries and calculate the various TL models’ per-

formance metrics for each group separately. This is

shown in Fig. 3 for TL with the complete dataset. BA

and IoU are plotted against the timestamp of the re-

spective group of images. The measured performance

metrics values vary over the entire timeseries, with

the tendency of decreasing performance towards later

timepoints where cells are more abundant and start

to accumulate and to overlap. Within error margins,

given as shaded areas in Fig. 3, TL models with nine

or more retrained layers yield the same performance

metrics while the two layer model performs slightly

worse.

Finally, we performed the same set of model train-

ing and evaluations for TL models trained on the par-

tial dataset. Note however, that the test split is iden-

tical to the complete dataset, i.e. it contains data

from one entire timeseries, including cell accumula-

tions and overlapping cells.

Table Tab. 2 lists the mean values for BA and IoU

over the timeseries for the models trained on the par-

tial training dataset. We observe the same trend as

in the case of training with the complete dataset: All

models yield the same average performance (within

the two sigma conﬁdence interval) with the model

Performance Review of Retraining and Transfer Learning of DeLTA2 for Image Segmentation for Pseudomonas Fluorescens SBW25

277

Figure 4: Balanced accuracy (left) and IoU (right) vs. image timestamp for transfer learning with varying number of retrained

layers for the partial MPB timeseries dataset. The shaded areas denote the min/max range over three independently trained

models. Horizontal lines next to y-axis indicate total mean values over the timeseries according to Tab. 1.

Table 2: Mean balanced accuracy and mean IoU for transfer

learning models trained on the partial MPB dataset and with

varying number of retrained layers. 0 retrained layers cor-

responds to the original DeLTA2 model. Error margins are

computed as 2 times the standard deviation over all test im-

ages and over three independently trained model replicates.

retrained layers mean BA mean IoU

0 0.85 0.65

2 0.86 ± 0.04 0.71 ±0.07

9 0.90 ± 0.04 0.76 ±0.06

18 0.90 ± 0.04 0.77 ±0.06

27 0.90 ± 0.04 0.78 ±0.05

36 0.90 ± 0.04 0.77 ±0.06

with only two retrained layers performing slightly

but insigniﬁcantly worse than the models with nine

or more retrained layers. Figure 4 shows the per-

formance metrics for images grouped by their times-

tamp. In this representation it becomes evident that

retraining only nine layers (25% of the DeLTA2 U–

Net) gives the same improvement as retraining the en-

tire model, while retraining only two layers performs

signiﬁcantly worse.

Interestingly, both BA and IoU values for TL with

the partial dataset coincide with the respective val-

ues for the complete dataset (Tab. 1). We compare

our transfer learning results for the partial and for

the complete training dataset in more detail in Fig. 5.

Within error bars, transfer learning with 10 retrained

layers yields the same validation metrics in both cases

at each timepoint in the series.

From these results we conclude that the test eval-

uation of TL models is not correlated with the com-

plexity of the training dataset.

6 DISCUSSION

Not surprising, we observe signiﬁcant improvement

in IoU and BA compared to the original DeLTA2

model when completely retraining the model on the

MPB dataset. We also conﬁrmed that TL improves

performance of the segmentation model.

As a rather novel result, we observed that the per-

formance gain with complete retraining and with TL

can be achieved with training data of reduced com-

plexity compared to the evaluation test data.

This last result paves the way towards a segmen-

tation workﬂow without the laborious manual label-

ing of training data. Manual labeling could be re-

placed by automated unsupervised labeling of the low

complexity images with only few isolated cells. In a

growth experiment, these would typically be the very

early images in the time series. These can efﬁciently

be segmented by e.g. thresholding or edge detection.

Subsequently, a U–Net could be trained on this au-

tomatically labeled data and ﬁnally be applied to the

full complexity dataset.

Looking out to future research, the present ap-

proach to evaluate various segmentation models and

their retraining should be extended to other Deep

Learning Segmentation implementations as well as to

traditional segmentation techniques. The latter play

BIOIMAGING 2024 - 11th International Conference on Bioimaging

278

Figure 5: Balanced Accuracy (left) and IoU (right) vs. image timestamp for two models trained on different datasets: Orange

lines for training with the complete training dataset, blue lines for training with the partial training dataset. Shaded areas

indicate the min-max interval from 3 model replicates. The test metrics are plotted against the timestep of the timelapse

microscopy test dataset.

a crucial role in the envisioned automated segmenta-

tion workﬂow and should be well understood in order

to achieve the best possible automated ground truth

labeling for the subsequent deep learning training.

ACKNOWLEDGEMENTS

We acknowledge Jana Grote for generating and pro-

viding the image data and Octavio Reyes-Matte for

stimulating discussion. BG and CFG acknowledge

generous support by the Max Planck Society.

REFERENCES

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.,

Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,

Askell, A., Agarwal, S., Herbert-Voss, A., Krueger,

G., Henighan, T., Child, R., Ramesh, A., Ziegler,

D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler,

E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner,

C., McCandlish, S., Radford, A., Sutskever, I., and

Amodei, D. (2020). Language Models are Few-Shot

Learners. Technical report. arXiv:2005.14165 [cs]

type: article.

Campello, V. M. et al. (2021). Multi-Centre, Multi-

Vendor and Multi-Disease Cardiac Segmentation: The

M&Ms Challenge. IEEE Transactions on Medical

Imaging, 40(12).

Chollet, F. (2018). Deep Learning With Python. Manning

Publications.

Cutler, K. J., Stringer, C., Lo, T. W., Rappez, L., Strous-

trup, N., Brook Peterson, S., Wiggins, P. A., and

Mougous, J. D. (2022). Omnipose: a high-precision

morphology-independent solution for bacterial cell

segmentation. Nature Methods, 19(11).

Ducret, A., Quardokus, E. M., and Brun, Y. V. (2016). Mi-

crobej, a tool for high throughput bacterial cell detec-

tion and quantitative analysis. Nature Microbiology,

1(7).

Fortmann-Grote, C. and other (2023). Raw Data for ”Per-

formance Review of Retraining and Transfer Learn-

ing of DeLTA 2.0 for Image Segmentation for Pseu-

domonas ﬂuorescens SBW25”.

Garc

ıa, V., Mollineda, R. A., and S

anchez, J. S. (2009). In-

dex of balanced accuracy: A performance measure for

skewed class distributions. In Araujo, H., Mendonc¸a,

A. M., Pinho, A. J., and Torres, M. I., editors, Pattern

Recognition and Image Analysis, volume 4 of Iberian

Conference on Pattern Recognition and Image Analy-

sis, IbPRIA. Springer, Berlin, Heidelberg.

Iman, M., Arabnia, H., and Rasheed, K. (2023). A review of

deep transfer learning and recent advancements. Tech-

nologies, 11.

Kim, H. E., Cosa-Linan, A., Santhanam, N., Jannesari, M.,

Maros, M. E., and Ganslandt, T. (2022). Transfer

learning for medical image classiﬁcation: a literature

review. BMC Medical Imaging, 22(1).

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-

ing. Nature, 521(7553).

Ma, J., He, Y., Li, F., Han, L., You, C., and Wang, B. (2023).

Segment anything in medical images. arXiv preprint

arXiv:2304.12306.

Minaee, S., Boykov, Y., Porikli, F., Plaza, A., Kehtarnavaz,

N., and Terzopoulos, D. (2020). Image segmenta-

Performance Review of Retraining and Transfer Learning of DeLTA2 for Image Segmentation for Pseudomonas Fluorescens SBW25

279

tion using deep learning: A survey. arXiv:2001.05566

[cs]. version: 1.

O’Connor, O. M., Alnahhas, R. N., Lugagne, J.-B., and

Dunlop, M. J. (2022). Delta 2.0: A deep learning

pipeline for quantifying single-cell spatial and tempo-

ral dynamics. PLOS Computational Biology, 18(1).

Panigrahi, S. et al. (2021). Misic, a general deep learning-

based method for the high-throughput cell segmenta-

tion of complex bacterial communities. eLife, 10.

Peng, H. (2008). Bioimage informatics: a new area of en-

gineering biology. Bioinformatics, 24(17).

Rainey, P. B. and Travisano, M. (1998). Adaptive radiation

in a heterogeneous environment. Nature, 394(6688).

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Lecture Notes in Computer Science.

Springer International Publishing.

Silby, M. W. et al. (2009). Genomic and genetic analyses of

diversity and plant interactions of pseudomonas ﬂuo-

rescens. Genome Biology, 10(5).

Stephens, Z. D. et al. (2015). Big data: Astronomical or

genomical? PLOS Biology, 13(7).

Stringer, C., Wang, T., Michaelos, M., and Pachitariu, M.

(2021). Cellpose: a generalist algorithm for cellular

segmentation. Nature Methods, 18.

Thompson, I. et al. (1993). Quantitative and qualitative sea-

sonal changes in the microbial community from the

phyllosphere of sugar beet (Beta vulgaris). Plant and

Soil, 150(2).

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L., and Polosukhin, I.

(2017). Attention is All you Need. In Advances in

Neural Information Processing Systems, volume 30.

Curran Associates, Inc.

Weiss, K., Khoshgoftaar, T. M., and Wang, D. (2016). A

survey of transfer learning. Journal of Big Data, 3(1).

Yu, X., Wang, J., Hong, Q., Teku, R., Wang, S.-H., and

Zhang, Y. (2022). Transfer learning for medical im-

ages analyses: A survey. Neurocomputing, 489.

BIOIMAGING 2024 - 11th International Conference on Bioimaging

280